A method for dynamic impedance calibration of high-frequency magnetic ring inductors based on convolutional neural networks

Abstract

This invention provides a dynamic impedance calibration method for high-frequency magnetic ring inductors based on convolutional neural networks, relating to the fields of radio frequency power electronics, switching power supplies, and electromagnetic compatibility design. The method includes: acquiring amplitude-frequency response, phase shift data, and ambient temperature of a high-frequency magnetic ring within the 0.1MHz–1GHz frequency band; after adaptive filtering, extracting amplitude-frequency timing features and phase shift / phase lag features using a dual-branch convolutional neural network; dynamically fusing these features using a bidirectional attention mechanism to generate a fused feature map; and finally outputting a dynamic calibration coefficient matrix to correct impedance measurements in real time. This invention solves the calibration accuracy degradation problem caused by device parasitic parameters and temperature drift in the high-frequency band (>400MHz) of traditional methods, achieving microsecond-level dynamic response over a wide temperature range. Furthermore, through multi-modal data fusion and noise immunity design, it significantly improves calibration robustness under high-frequency nonlinear conditions.

Description

Technical Field

[0001] This invention relates to the fields of radio frequency power electronics, switching power supplies and electromagnetic compatibility design technology, and in particular to a method for dynamic impedance calibration of high-frequency magnetic ring inductors based on convolutional neural networks. Background Technology

[0002] In the fields of RF power electronics, switching power supplies, and electromagnetic compatibility design, the impedance characteristics of high-frequency magnetic toggle inductors have a decisive impact on system performance. With operating frequencies expanding to the GHz level (such as in 5G base station power amplifiers and GaN fast-charging circuits), traditional impedance calibration methods face severe challenges.

[0003] High-frequency calibration accuracy deteriorates. Existing technologies rely on impedance bridge networks constructed from standard capacitors / resistors, which exhibit significant errors in the >400MHz frequency band. Parasitic parameters of standard devices (such as lead inductance and distributed capacitance) produce non-ideal effects as frequency increases, leading to large calibration deviations. More seriously, temperature drift causes devices to become further inaccurate under high-temperature conditions, failing to meet the requirements of wide-temperature-range applications.

[0004] Lack of dynamic response capability. The impedance of the magnetic ring changes dynamically with load current, temperature, and aging, but traditional methods use static calibration coefficients. For example, in LLC resonant converters, the impedance of the integrated magnetic inductor decreases due to magnetic saturation under full load, and existing calibration schemes cannot track this change in real time, resulting in energy efficiency loss.

[0005] Insufficient nonlinear characterization. Magnetic rings exhibit strong nonlinear characteristics under high-frequency operating conditions: eddy current losses in ferrite cores increase exponentially at frequencies above 300MHz; interlayer parasitic capacitances cause resonance peak shifts at frequencies above 500MHz; and the frequency-varying permeability characteristics of nanocrystalline magnetic rings lead to phase lag. Existing technologies rely solely on the amplitude-frequency response as a single parameter, failing to capture the coupling relationship between phase lag and nonlinear losses.

[0006] Sensitive to noise interference. The dV / dt noise and RF radiation from the switching power supply degrade the signal-to-noise ratio of the measurement signal. Traditional low-pass filters suppress noise but also filter out high-frequency phase information, resulting in large phase calibration errors.

[0007] Calibration efficiency is low. Manual balancing calibration based on impedance bridges requires iterative adjustments to standard component values, resulting in a long calibration time per cycle. In mass production scenarios, this process becomes a bottleneck for production capacity, and manual operation introduces random errors. Summary of the Invention

[0008] To address the technical problems of degraded high-frequency calibration accuracy, lack of dynamic response capability, insufficient nonlinear characterization, sensitivity to noise interference, and low calibration efficiency in existing technologies, this invention provides a high-frequency magnetic ring inductor dynamic impedance calibration method based on convolutional neural networks.

[0009] The technical solution provided by this invention is as follows:

[0010] This invention provides a method for dynamic impedance calibration of high-frequency magnetic ring inductors based on convolutional neural networks, comprising:

[0011] S1. Collect multi-dimensional time-series data of the high-frequency magnetic ring in the 0.1MHz–1GHz frequency band, including the amplitude frequency response sequence A(f,t), the phase shift data sequence θ(f,t), and the ambient temperature T;

[0012] S2. Perform adaptive filtering on the multidimensional time-series data to suppress electromagnetic noise interference;

[0013] S3. Construct a two-branch convolutional neural network, where branch 1 processes the temporal characteristics of the amplitude-frequency response sequence and branch 2 processes the phase lag characteristics of the phase-shifted data sequence.

[0014] S4. Dynamically fuse the feature vectors output from branch 1 and branch 2 using a bidirectional attention mechanism to generate a fused feature map;

[0015] S5. Calculate the dynamic frequency resolution parameter δ f and phase consistency index η θ Optimize the feature extraction range;

[0016] S6. Input the fused feature map into the fully connected regression layer, and output the dynamic calibration coefficient matrix K. cal =[k g ,k φ ], where k g k is the gain compensation coefficient. φ This refers to the phase compensation coefficient;

[0017] S7. Correct the measured magnetic ring impedance value Z in real time according to the calibration coefficient matrix. meas Generate calibrated impedance

[0018] Furthermore, S2 further includes:

[0019] S201. Calculate the signal-to-noise ratio (SNR) of the amplitude-frequency response sequence A(f,t);

[0020] S202. Dynamically determine the length of the moving average filter window based on the signal-to-noise ratio:

[0021]

[0022] Where f s Where Δf is the sampling frequency. max The maximum frequency interval is SNR, which is the signal-to-noise ratio.

[0023] S203. Perform a moving average filter on A(f,t) using a window length L;

[0024] S204. Perform one-dimensional convolution filtering on the phase-shifted data sequence θ(f,t), with the convolution kernel weights based on the phase standard deviation σ. θ generate.

[0025] Furthermore, the processing of branch 2 in S3 specifically includes:

[0026] S301, Obtain the magnetic ring conduction rise time t rise The results were obtained through impulse response testing.

[0027] S302. Construct a phase differential convolutional layer, the operation of which adopts a phase gradient model:

[0028]

[0029] Where θ(f,t) is the phase-shifted data sequence, f is the frequency, and t is the time. rise This refers to the rise time of the magnetic ring during conduction.

[0030] S303. Input the phase-shifted data sequence into the phase differential convolutional layer to extract the phase lag features.

[0031] Furthermore, the calculation of the dynamic frequency resolution parameter in S5 includes:

[0032] S501, Calculate f for each frequency point i rate of change of amplitude

[0033] S502, Synthetic Dynamic Frequency Resolution Parameters:

[0034]

[0035] Where N is the number of frequency points, A(f i ) is the frequency point f i At the amplitude, σ noise For the noise standard deviation, A avg This represents the average amplitude.

[0036] Furthermore, the calculation of the phase consistency index in S5 includes:

[0037] S503, Calculate the phase deviation of each sampling point in The average of the phases;

[0038] S504, Generate frequency-weighted phase consistency index:

[0039]

[0040] Where M is the number of sampling points, θ k Let k be the phase value at the kth sampling point. f is the phase mean. k f is the frequency of the kth sampling point. max α is the maximum sampling frequency, and α is the adjustable frequency weighting factor.

[0041] Furthermore, the execution process of S4 specifically includes:

[0042] S401, the feature vector h of branch 1 i and the eigenvector h of branch 2 j Concatenate into a joint vector [h] i h j ];

[0043] S402. Calculate the energy function using an improved Bahdanau attention mechanism:

[0044]

[0045] Where h i Output the feature vector for branch 1, h j Output feature vector W for branch 2 a b is a trainable weight matrix a v is a trainable bias vector. a This is the weight vector;

[0046] S403, Normalize the energy function to generate attention weights;

[0047] S404. Fuse feature vectors based on attention weights.

[0048] Furthermore, S6 further includes:

[0049] S601. Principal component analysis is used to reduce the dimensionality of the fused feature map;

[0050] S602, Based on dynamic frequency resolution parameter δ f Filter key frequency band feature components;

[0051] S603. Input the filtered features into the random forest ensemble algorithm and output the calibration coefficient matrix.

[0052] Furthermore, model anti-interference training is performed before S1, specifically including:

[0053] S001. Generate a synthetic training dataset containing white noise and impulse interference;

[0054] S002. Construct a composite loss function to train the CNN model:

[0055]

[0056] Where λ1 and λ2 are the loss weight coefficients, Z cal For the calibrated impedance, Z ref As the reference impedance, Represents the frequency domain gradient operator;

[0057] S003. Optimize model parameters using an adversarial training strategy.

[0058] Furthermore, it also includes S8:

[0059] S801, Real-time calculation of calibration error ∈=|Z cal -Z true | / Z true ;

[0060] S802. When ∈>5%, the online update mechanism is triggered:

[0061] S8021. Collect current operating condition data as incremental training samples;

[0062] S8022, Fine-tune the weights of the fully connected layers in the CNN model.

[0063] Furthermore, the operations performed by the dual-branch CNN in S3 specifically include:

[0064] S310, Output F from branch 1 branch1 and branch 2 output F branch2 Input residual module generates F res ;

[0065] S311. Feature fusion via attention gating mechanism:

[0066] F out =ReLU(F res +β·ATT(F branch1 ,F branch2 ))

[0067] Where F res For the residual module output characteristics, β is determined by the phase consistency index η. θ Dynamic adjustment, ATT(·) is the attention operation.

[0068] The beneficial effects of the technical solution provided by this invention include at least the following:

[0069] (1) In this invention, a convolutional neural network is constructed to directly learn the amplitude-phase frequency timing characteristics of high-frequency magnetic ring inductors in the 0.1MHz–1GHz frequency band. Combined with the dynamic calibration coefficients output by a fully connected regression layer, microsecond-level online impedance compensation is achieved. This scheme eliminates the dependence on physical standard devices and effectively solves the problem of calibration accuracy degradation caused by device parasitic parameters and temperature drift in the high-frequency band (>400MHz), ensuring high-precision dynamic response capability in a wide temperature range (-40℃~125℃).

[0070] (2) In this invention, a dual-branch convolutional neural network architecture is designed to process the amplitude frequency data time-series vector and the phase frequency data time-series vector respectively, and the two types of feature vectors are dynamically fused through a bidirectional attention mechanism. This scheme enhances the characterization ability of high-frequency nonlinear characteristics (such as core eddy current loss and parasitic capacitance coupling), solves the problem that traditional methods cannot capture the coupling relationship between phase lag and nonlinear loss by relying on a single impedance parameter, and significantly improves the calibration accuracy under complex working conditions.

[0071] (3) In this invention, an adaptive filtering layer is embedded in the front end of the CNN, combining moving average filtering with one-dimensional convolution to suppress non-periodic noise, and synthetic noise data (such as white noise and impulse interference) is injected during the training phase to improve the model's generalization ability. This scheme effectively preserves phase information under high-frequency electromagnetic interference environment, solves the problem of phase information loss caused by traditional low-pass filtering, and enhances the model's robustness to unseen interference types through adversarial training strategy, ensuring the stability of high-precision calibration. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 A flowchart illustrating a high-frequency magnetic ring inductor dynamic impedance calibration method based on a convolutional neural network, provided in an embodiment of the present invention;

[0074] Figure 2 This is a schematic diagram of the filtering process in a high-frequency magnetic ring inductor dynamic impedance calibration method based on a convolutional neural network, provided in an embodiment of the present invention. Detailed Implementation

[0075] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0076] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0077] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent.

[0078] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0079] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0080] Reference manual attached Figure 1 The diagram shows a flowchart of a high-frequency magnetic ring inductor dynamic impedance calibration method based on a convolutional neural network provided by an embodiment of the present invention.

[0081] This invention provides a method for dynamic impedance calibration of high-frequency magnetic ring inductors based on convolutional neural networks. The processing flow may include the following steps:

[0082] S1. Collect multi-dimensional time-series data of the high-frequency magnetic ring in the 0.1MHz–1GHz frequency band, including the amplitude frequency response sequence A(f,t), the phase shift data sequence θ(f,t), and the ambient temperature T.

[0083] In one possible implementation, the amplitude-frequency response sequence A(f,t) and phase-shift data sequence θ(f,t) of the magnetic ring are acquired using a vector network analyzer (VNA), with a sampling frequency f. s The GHz frequency must be no less than 2 GHz to satisfy the Nyquist sampling theorem. The ambient temperature T is monitored in real time by a high-precision thermocouple mounted on the surface of the magnetic ring, and the temperature data is transmitted via I... 2 The data is transmitted to the processing unit via the C-bus. To cover the full frequency band, a linear frequency sweep mode is adopted, with a sampling duration of ≥10 signal cycles for each frequency point to ensure data integrity.

[0084] It should be noted that data acquisition must be carried out in an electromagnetically shielded room with a shielding effectiveness of ≥80dB (10kHz-1GHz) to avoid external electromagnetic interference; the test platform adopts a shockproof design, and the temperature is controlled at 25±2℃ and the humidity ≤60%RH to reduce the impact of environmental fluctuations.

[0085] It should be noted that when acquiring the amplitude-frequency response sequence A(f,t), a 50Ω terminating resistor needs to be connected in parallel across the magnetic ring to suppress signal reflection. The test fixture adopts a PCB microstrip line structure with a characteristic impedance of 50Ω. A 10Ω current-limiting resistor is connected in series at the input to suppress overshoot, and an RC absorption network is connected in parallel at the output to eliminate high-frequency oscillations, ensuring that the reflection coefficient is ≤-20dB in the 0.1MHz-1GHz frequency band. The sampling rate of the ambient temperature T should be synchronized with the electromagnetic data, with a sampling interval ≤10ms. Temperature drift compensation uses a third-order polynomial fitting:

[0086] T comp =T raw -(c0+c1t+c2t 2 +c3t 3 )

[0087] The coefficients c0 to c3 were obtained through a constant temperature chamber calibration experiment.

[0088] After completing the acquisition of multi-dimensional time-series data, the amplitude-frequency response and phase shift data may contain glitches or distortions due to the susceptibility of high-frequency magnetic rings to interference from electromagnetic radiation and environmental noise in the 0.1MHz–1GHz frequency band. If these noises are not processed, they will directly affect the accuracy of subsequent feature extraction. Therefore, adaptive filtering is needed to clean the data and provide a reliable input foundation for the neural network.

[0089] S2. Perform adaptive filtering on multidimensional time-series data to suppress electromagnetic noise interference.

[0090] In one possible implementation, such as Figure 2 As shown, S2 further includes:

[0091] S201. Calculate the signal-to-noise ratio (SNR) of the amplitude-frequency response sequence A(f,t).

[0092] Specifically, this is achieved through the power spectral density method: the signal frequency band is divided into sub-bands, the ratio of signal power to noise power in each sub-band is calculated, and the geometric mean is taken as the global SNR.

[0093] Furthermore, the subbands are divided in logarithmic intervals: 0.1-1MHz is divided into 1 subband every 100kHz, 1-10MHz into 1 subband every 1MHz, and 10-1000MHz into 1 subband every 10MHz, ensuring that each subband contains a sufficient signal period.

[0094] S202. Dynamically determine the length of the moving average filter window based on the signal-to-noise ratio:

[0095]

[0096] Where f s Where Δf is the sampling frequency. max The maximum frequency interval is given by SNR, which is the signal-to-noise ratio. This formula ensures that the window is increased to smooth noise when the signal-to-noise ratio is high, and the window is reduced to preserve details when the signal-to-noise ratio is low.

[0097] S203. A moving average filter is applied to A(f,t) using a window length L. The filtered data is stored in a circular buffer, and the buffer depth is proportional to the number of frequency points.

[0098] S204. Perform one-dimensional convolution filtering on the phase-shifted data sequence θ(f,t), with the convolution kernel weights based on the phase standard deviation σ. θ Specifically, the standard deviation of the phase-shifted data within the current sliding window is calculated, and the center weight of the convolution kernel is set to 1 / σ. θ The edge weights decay according to a Gaussian distribution.

[0099] It should be noted that the sliding window adopts a non-overlapping method, and the window step size is consistent with the frequency point interval. During the filtering process, if the output difference between adjacent windows exceeds twice the phase standard deviation, a weighted average transition is enabled to avoid filtering artifacts.

[0100] After adaptive filtering, noise interference in the data is effectively suppressed, and the temporal variation law of amplitude-frequency response and the phase lag characteristics of phase-shifted data are highlighted. Since amplitude-frequency response and phase characteristics reflect the amplitude and phase dimensions of magnetic ring impedance, respectively, and the correlation between the two differs in different frequency bands, a dual-branch convolutional neural network is designed to extract the two types of features separately to more comprehensively capture the dynamic characteristics of impedance.

[0101] S3. Construct a two-branch convolutional neural network (CNN), where branch 1 processes the temporal features of the amplitude-frequency response sequence: a 5-layer one-dimensional convolutional structure is adopted, the kernel width of the first layer is 3, the stride is 1, the activation function is ReLU (Rectified Linear Unit), and the kernel width of subsequent layers increases to extract long-term dependencies.

[0102] Branch 2 processes the phase hysteresis characteristics of phase-shifted data sequences:

[0103] In one possible implementation, the processing of branch 2 in S3 specifically includes:

[0104] S301. Obtain the magnetic ring conduction rise time t through pulse response testing.rise The current was measured by a pulse response test, specifically by applying a step voltage to the magnetic ring and recording the time required for the current to rise to 90% of the rated value.

[0105] S302. Construct a phase differential convolutional layer, the operation of which adopts a phase gradient model:

[0106]

[0107] Where θ(f,t) is the phase-shifted data sequence, f is the frequency, and t is the time. rise The phase differential convolutional layer is implemented through differentiable programming to represent the rise time of the magnetic ring conduction, and outputs a phase change rate feature map.

[0108] It should be noted that the phase differential convolutional layer is implemented using an automatic differentiation framework (such as PyTorch) to avoid errors from manual differentiation. The computational graph construction process during forward propagation is as follows:

[0109] 1. Register the phase-shifted data sequence θ(f,t) as a differentiable tensor;

[0110] 2. Call torch.autograd.grad() to calculate.

[0111] 3. With 1 / t rise Perform element-wise multiplication.

[0112] S303. Input the phase-shifted data sequence into the phase differential convolutional layer to extract phase lag features, and keep the output dimension consistent with branch 1.

[0113] It should be noted that branch 1 gradually increases the number of channels to 512 through 5 layers of convolution, while branch 2 matches the dimension through 3 layers of convolution (number of channels [64,256,512]) after phase micro-layering. Finally, both are compressed into 512-dimensional feature vectors through global average pooling.

[0114] It should be noted that the architecture parameters of a two-branch CNN are as follows:

[0115] Branch 1: 5 layers of one-dimensional convolution, with kernel width sequence [3,5,7,9,11] and number of channels [32,64,128,256,512];

[0116] Branch 2: After phase micro-layering, there are 3 layers of convolution, with the kernel width fixed at 3;

[0117] Batch normalization and ReLU activation for each layer.

[0118] Furthermore, the output features of branch 1 and branch 2 are adjusted to the same size through interpolation, where linear interpolation is used in the time dimension and sinc interpolation is used in the frequency dimension to reduce aliasing; the feature alignment error needs to be controlled within ±1 sampling point.

[0119] In one possible implementation, the operations performed by the dual-branch CNN of S3 specifically include:

[0120] S310, Output F from branch 1 branch1 and branch 2 output F branch2 Input residual module generates F res :

[0121] F res =ReLU(Conv1D(F branch1 )+Conv1D(F branch2 ))

[0122] The kernel width is 1, used to adjust the channel dimension.

[0123] S311. Feature fusion via attention gating mechanism:

[0124] F out =ReLU(F res +β·ATT(F branch1 ,F branch2 ))

[0125] Where F res For the residual module output characteristics, β is determined by the phase consistency index η. θ Dynamic adjustment, ATT(·) is the attention operation.

[0126] After the dual-branch network extracts the time-series features of the amplitude-frequency response and the phase lag features of the phase shift data respectively, the importance of the two types of features changes dynamically in different frequency bands (for example, the phase features in the high-frequency band have a more significant impact on impedance calibration). To achieve dynamic weighted fusion of features, a bidirectional attention mechanism is adopted, enabling the model to adaptively focus on key feature components and improve the ability of the fused feature map to represent impedance changes.

[0127] S4. Dynamically fuse the feature vectors output from branch 1 and branch 2 through a bidirectional attention mechanism to generate a fused feature map.

[0128] In one possible implementation, the execution process of S4 specifically includes:

[0129] S401, the feature vector h of branch 1 i and the eigenvector h of branch 2 j Concatenate into a joint vector [h] i h j ].

[0130] S402. Calculate the energy function using an improved Bahdanau attention mechanism:

[0131]

[0132] Where h i Output a feature vector (dimension 128) for branch 1, h j Output a feature vector (dimension 128) for branch 2, W a The weight matrix is ​​a 128×256 trainable weight matrix, b a v is a 256-dimensional trainable bias vector. a For a 256-dimensional trainable attention weight vector, η θ This is the phase consistency index.

[0133] It should be noted that when the attention weight is concentrated in the high frequency band (>500MHz), the feature weight ratio of branch 2 is automatically increased; when the frequency band (<100MHz), the contribution of branch 1 is enhanced, and the weight allocation dynamically adapts to the frequency band characteristics.

[0134] S403, regarding the energy function e ij Softmax normalization is performed to generate attention weights;

[0135] S404. Generate a 256-dimensional fused feature map by weighting and summing the feature vectors according to the attention weights.

[0136] It should be noted that the dimensionality compression method for fused feature maps is as follows:

[0137] 1. When the number of feature map channels > 512, use 1×1 convolution for dimensionality reduction;

[0138] 2. The max pooling layer step size is set to 2 to preserve peak features;

[0139] 3. Channel attention mechanism weight calculation:

[0140] w c =σ(MLP(AvgPool(F)+MLP(MaxPool(F)))

[0141] Where w c σ is the channel attention weight vector (dimension equal to the number of channels), MLP stands for Multilayer Perceptron, which contains two fully connected layers, AvgPool(·) represents the global average pooling operation, MaxPool(·) represents the global max pooling operation, and F represents the input feature map (dimension [C,H,W]).

[0142] The fused feature map integrates key information about amplitude and phase, but the feature extraction range needs to be dynamically adjusted according to the actual characteristics of the magnetic ring—for example, denser feature sampling is needed in frequency bands with drastic amplitude changes, while the feature range needs to be narrowed in frequency bands with poor phase stability. Therefore, the dynamic frequency resolution parameter δ is calculated. f and phase consistency index η θ This provides a quantitative basis for optimizing the range of feature extraction, ensuring that the calculation of subsequent calibration coefficients is based on the most representative features.

[0143] S5. Calculate the dynamic frequency resolution parameter δ f and phase consistency index η θ Optimize the feature extraction range.

[0144] In one possible implementation, the calculation of the dynamic frequency resolution parameter in S5 includes:

[0145] S501, Calculate f for each frequency point i rate of change of amplitude

[0146] S502, Synthetic Dynamic Frequency Resolution Parameters:

[0147]

[0148] Where N is the number of frequency points, A(f i ) is the frequency point f i At the amplitude, σ noise For the noise standard deviation, A avg This represents the average amplitude. The exponential term achieves adaptive noise suppression, when σ... noise >A avg Significantly reduced δ f value.

[0149] It should be noted that, according to the dynamic frequency resolution parameter δ f Determine the frequency step size for feature extraction: δ f When the value is large, decrease the step size (dense sampling), δ f Increase the step size when the value is small (sparse sampling); phase consistency index η θ Used to filter phase-stable frequency bands, with priority given to retaining η. θ Frequency band characteristics >0.8.

[0150] It should be noted that, in order to obtain the σ parameter in the dynamic frequency resolution parameter... noise ,implement:

[0151] 1. Select a frequency band of 30-40MHz (this band has no signal components);

[0152] 2. Calculate the standard deviation of the amplitude data for this frequency band:

[0153]

[0154] Where σ noise The noise standard deviation is represented in dB, K represents the number of sampling points in the 30-40MHz frequency band, and A represents the noise standard deviation (in dB). k Let μ represent the amplitude (in dB) at the k-th sampling point, and μ represent the average amplitude of that frequency band. (Unit: dB)

[0155] In one possible implementation, the calculation of the phase coherence index in S5 includes:

[0156] S503, Calculate the phase deviation of each sampling point in The average of the phases;

[0157] S504, Generate frequency-weighted phase consistency index:

[0158]

[0159] Where M is the number of sampling points, θ k Let k be the phase value at the kth sampling point. f is the phase mean. k f is the frequency of the kth sampling point. max The maximum sampling frequency is α, which is an adjustable frequency weighting factor (values ​​range from 0.5 to 2.0). Frequency weighting term [f] k / f max ] α Strengthen the phase stability weight in the high-frequency band.

[0160] It should be noted that the α value is dynamically adjusted according to the type of magnetic ring material: 1.5 for ferrite magnetic rings (poor high-frequency phase stability), 0.8 for nanocrystalline magnetic rings (high phase consistency), and 1.0 for manganese-zinc ferrite (a compromise).

[0161] Based on δ f and η θ After optimizing the feature extraction range, the fused feature map retains only the key components strongly correlated with impedance calibration. At this point, the features are input into a fully connected regression layer. Combined with principal component analysis dimensionality reduction and a random forest ensemble algorithm, the relationship between the features and calibration coefficients can be accurately mapped. The output is a dynamic calibration coefficient matrix that can simultaneously compensate for amplitude and phase errors, providing a quantitative basis for correcting impedance measurements.

[0162] S6. Input the fused feature map into the fully connected regression layer, and output the dynamic calibration coefficient matrix K. cal =[k g ,k φ ], where kg k is the gain compensation coefficient. φ This is the phase compensation coefficient.

[0163] In one possible implementation, S6 further includes:

[0164] S601. Principal Component Analysis (PCA) is used to reduce the dimensionality of the fused feature map, retaining the principal components corresponding to 95% of the energy.

[0165] S602, Based on dynamic frequency resolution parameter δ f Screening key frequency band feature components: Calculate the variance of each frequency band feature component, and retain those with variance greater than δ. f The amount.

[0166] S603. Input the filtered features into the Random Forest Ensemble Algorithm. The random forest contains 100 decision trees, each with a maximum depth of 10. The splitting criterion is the mean squared error, and the output is a calibration coefficient matrix K. cal =[k g ,k φ ].

[0167] It should be noted that the optimization strategy of the random forest ensemble algorithm is as follows:

[0168] 1. Feature importance is assessed using the Gini index:

[0169]

[0170] Where C is the number of categories, p i The proportion of categories in the node;

[0171] 2. The feature sampling ratio during decision tree splitting is set to... (d is the feature dimension);

[0172] 3. Set the minimum number of samples per node to 5 to prevent overfitting.

[0173] S7. Correct the measured magnetic ring impedance value Z in real time according to the calibration coefficient matrix. meas Generate calibrated impedance Z meas k was measured directly by an impedance analyzer. g and k φ These are applied to amplitude and phase compensation, respectively. The compensation results are accelerated to microsecond-level latency using FPGA (Field Programmable Gate Array) hardware.

[0174] It should be noted that the complex number operations for phase compensation are expanded using Euler's formula as follows:

[0175] e jkφ =cos(kφ)+jsin(kφ)

[0176] The real part corresponds to the amplitude correction, the imaginary part corresponds to the phase correction, and the calculation precision is retained to 6 decimal places.

[0177] It should be noted that in the FPGA hardware acceleration implementation, the calibration calculation pipeline is designed as follows: sampling → buffering → CNN inference → coefficient application → output. The critical timing inter-layer delay of the CNN is ≤50ns, the coefficient multiplier uses a DSP48E1 unit, and the phase compensation uses the CORDIC algorithm.

[0178] Before formally executing the data acquisition and calibration process, it is necessary to improve the robustness of the model through anti-interference training. Since magnetic rings may face complex electromagnetic environments such as white noise and impulse interference in practical applications, by synthesizing training datasets containing interference and training with composite loss functions, the model can maintain stable feature extraction and calibration capabilities even when facing unseen types of interference during subsequent actual data processing.

[0179] In one possible implementation, model robustness training is performed before S1, specifically including:

[0180] S001. Generate a synthetic training dataset containing white noise and impulse interference. Specifically, to generate the synthetic training dataset, white noise (power 10%-50% of signal power), impulse interference (pulse width ≤ 100ns), and frequency aliasing (offset ±5%) are injected into the clean data. The clean data is derived from measurements of a standard magnetic ring under ideal conditions: acquired under constant temperature (25℃), shielding (≥100dB), and no-load conditions, ensuring a signal-to-noise ratio ≥40dB.

[0181] S002. Construct a composite loss function to train the CNN model:

[0182]

[0183] Where λ1 and λ2 are the loss weight coefficients, Z cal For the calibrated impedance, Z ref As the reference impedance, This represents the frequency domain gradient operator (implemented via the derivative of the fast Fourier transform).

[0184] It should be noted that the initial values ​​of the loss weight coefficients λ1 and λ2 are set to 1:1, and are dynamically adjusted according to the calibration error during training: if the phase error accounts for more than 50%, λ2 will automatically increase to 1.5 times that of λ1, and vice versa.

[0185] S003. Optimize model parameters using adversarial training strategy: Limit the perturbation amplitude of generated adversarial examples to within ±3dB, and update model weights using the FGSM (Fast Gradient Sign Method) algorithm.

[0186] The calibrated impedance Z obtained from the calibration coefficient matrix cal This already meets the accuracy requirements under most operating conditions. However, the impedance characteristics of the magnetic ring may slowly degrade with long-term use, temperature drift, or load changes, leading to a gradual increase in calibration error. Therefore, an online update mechanism is designed. When the calibration error exceeds the threshold in real time, the model is fine-tuned through incremental training to ensure that the calibration accuracy remains stable throughout the entire lifespan of the magnetic ring.

[0187] In one possible implementation, S8 is also included:

[0188] S801, Real-time calculation of calibration error ∈=|Z cal -Z true | / Z true Z true Calibration is performed using the four-terminal method.

[0189] S802. When ∈>5%, the online update mechanism is triggered:

[0190] S8021. Collect current operating condition data (including parameters such as temperature T and load current) as incremental training samples.

[0191] It should be noted that the operating data includes: load current (0-5A, sampling rate 1kHz), magnetic ring surface temperature distribution (3-point temperature measurement, error ±0.5℃), ambient humidity (30%-70%RH) and power supply voltage fluctuation (±5% of rated value).

[0192] S8022, Fine-tuning the weights of fully connected layers in a CNN model: Freeze the weights of the CNN convolutional layers and fine-tune only the weights of the fully connected layers, setting the learning rate to 10% of the initial value.

[0193] It should be noted that incremental training samples are stored using a sliding window, retaining only the operating data from the most recent 3 months, while older data is discarded according to the 'first-in, first-out' principle; model parameters are automatically backed up after each update, retaining the most recent 5 versions to support the rollback mechanism.

[0194] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0195] (1) In this invention, a convolutional neural network is constructed to directly learn the amplitude-phase frequency timing characteristics of high-frequency magnetic ring inductors in the 0.1MHz–1GHz frequency band. Combined with the dynamic calibration coefficients output by a fully connected regression layer, microsecond-level online impedance compensation is achieved. This scheme eliminates the dependence on physical standard devices and effectively solves the problem of calibration accuracy degradation caused by device parasitic parameters and temperature drift in the high-frequency band (>400MHz), ensuring high-precision dynamic response capability in a wide temperature range (-40℃~125℃).

[0196] (2) In this invention, a dual-branch convolutional neural network architecture is designed to process the amplitude frequency data time-series vector and the phase frequency data time-series vector respectively, and the two types of feature vectors are dynamically fused through a bidirectional attention mechanism. This scheme enhances the characterization ability of high-frequency nonlinear characteristics (such as core eddy current loss and parasitic capacitance coupling), solves the problem that traditional methods cannot capture the coupling relationship between phase lag and nonlinear loss by relying on a single impedance parameter, and significantly improves the calibration accuracy under complex working conditions.

[0197] (3) In this invention, an adaptive filtering layer is embedded in the front end of the CNN, combining moving average filtering with one-dimensional convolution to suppress non-periodic noise, and synthetic noise data (such as white noise and impulse interference) is injected during the training phase to improve the model's generalization ability. This scheme effectively preserves phase information under high-frequency electromagnetic interference environment, solves the problem of phase information loss caused by traditional low-pass filtering, and enhances the model's robustness to unseen interference types through adversarial training strategy, ensuring the stability of high-precision calibration.

[0198] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0199] The following points need to be explained:

[0200] (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.

[0201] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the present invention; that is, these drawings are not drawn to actual scale. It is understood that when an element such as a layer, film, region, or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element, or there may be intermediate elements.

[0202] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0203] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A high-frequency magnetic ring inductance dynamic impedance calibration method based on a convolutional neural network, characterized by, Comprise: S1, collecting multi-dimensional time series data of the high-frequency magnetic ring in the 0.1 MHz-1 GHz frequency band, including amplitude-frequency response sequence , phase shift data sequence , and ambient temperature ; S2, adaptive filtering processing of the multi-dimensional time series data, inhibiting electromagnetic noise interference; S3, constructing a double-branch convolutional neural network, wherein branch 1 processes the time series characteristics of the amplitude-frequency response sequence, and branch 2 processes the phase lag characteristics of the phase shift data sequence; S4, dynamically fusing the feature vectors output by branch 1 and branch 2 through a bidirectional attention mechanism to generate a fused feature map; S5, calculating a dynamic frequency resolution parameter and a phase consistency index , optimizing a feature extraction range; S6, input the fusion feature map into a fully connected regression layer to output a dynamic calibration coefficient matrix wherein is a gain compensation coefficient, is a phase compensation coefficient; S7、according to the calibration coefficient matrix, real-time correction of the magnetic ring impedance measurement value , generating the calibrated impedance ; The execution process of S4 specifically comprises: S401, concatenate the feature vector of branch 1 and the feature vector of branch 2 into a joint vector ; S402, calculating an energy function using an improved Bahdanau attention mechanism: ； wherein outputs a feature vector for branch 1, outputs a feature vector for branch 2, is a trainable weight matrix, is a trainable bias vector, is a weight vector; S403, normalizing the energy function to generate attention weights; S404, fusing feature vectors according to the attention weights.

2. The high-frequency magnetic loop inductance dynamic impedance calibration method based on a convolutional neural network according to claim 1, characterized in that, The S2 further comprises: S201、compute a sequence of amplitude-frequency responses signal-to-noise ratio SNR; S202, dynamically determining the length of the moving average filter window according to the signal-to-noise ratio: ； wherein is the sampling frequency, is the maximum frequency interval, is the signal-to-noise ratio; S203, employing a window length to performing a moving average filter; S204, phase shifting the data sequence performing one-dimensional convolution filtering, the convolution kernel weights being based on the phase standard deviation generation.

3. The method of claim 1, wherein the method is based on a convolutional neural network. The processing of branch 2 in S3 specifically comprises: S301, acquire the magnetic ring conduction rising time measured by impulse response test; S302, constructing a phase differential convolution layer, and the operation process thereof adopts a phase gradient model: ； wherein is a phase-shifted data sequence, is a frequency, is a time, is a magnetic ring conduction rise time; S303, inputting the phase shift data sequence into the phase differential convolution layer to extract phase lag features.

4. The high-frequency magnetic loop inductance dynamic impedance calibration method based on a convolutional neural network according to claim 1, characterized in that, The calculation of the dynamic frequency resolution parameter in S5 comprises: S501、Calculate the amplitude change rate of each frequency point S501、Calculate the amplitude change rate of each frequency point ; S502, synthesizing a dynamic frequency resolution parameter: ； wherein is the number of frequency points, is the frequency point is the amplitude at, is the standard deviation of the noise, is the average amplitude.

5. The method of claim 1, wherein the method is based on a convolutional neural network. The calculation of the phase consistency index in S5 comprises: S503, calculate phase deviation of each sampling point wherein is the phase mean; S504, generating a frequency-weighted phase consistency index: ； wherein is the number of samples, is the first sample point phase value, is the phase mean, is the first sample point frequency, is the maximum sampling frequency, is the adjustable frequency weight factor.

6. The method of claim 4, wherein the method is based on a convolutional neural network. The S6 further comprises: S601, using principal component analysis to reduce the dimensionality of the fused feature map; S602, based on the dynamic frequency resolution parameter Screening key band characteristic components; S603, inputting the filtered features into a random forest ensemble algorithm to output a calibration coefficient matrix.

7. The method of claim 1, wherein the method is based on a convolutional neural network. The model anti-interference training is performed before S1, specifically comprising: S001, generating a synthetic training data set containing white noise and impulse interference; S002, constructing a composite loss function to train the CNN model: ； wherein , is a loss weight coefficient, is a calibrated impedance, is a reference impedance, denotes a frequency domain gradient operator; S003, using an adversarial training strategy to optimize the model parameters.

8. The high-frequency magnetic loop inductance dynamic impedance calibration method based on a convolutional neural network according to claim 1, characterized in that, Further comprising S8: S801, calculate calibration error in real time ; S802、When the online update mechanism is triggered: S8021, collecting current working condition data as incremental training samples; S8022, fine-tuning the CNN model full connection layer weights.

9. The high-frequency magnetic loop inductance dynamic impedance calibration method based on a convolutional neural network according to claim 5, characterized in that, The operation performed by the double-branch CNN of S3 specifically comprises: S310, outputting the branch 1 and the branch 2 input residual module generation ; S311, fusing features through an attention gate mechanism: ； wherein is a residual module output feature, by a phase consistency index is dynamically adjusted, is an attention operation.

Images