DC-DC converter output end DC bus capacitance parameter identification method based on DWT and IBS-CollNet

Abstract

The invention provides a DC-DC converter output end DC bus capacitor parameter identification method based on DWT and IBS-CollNet, and the method comprises the steps: collecting the output voltages of a DC bus capacitor at different aging degrees, and forming a voltage ripple signal data set; carrying out DWT conversion on the voltage ripple signal data to obtain frequency band information, and carrying out importance sorting on frequency bands; determining the number of frequency bands to be screened based on a threshold segmentation algorithm; searching a continuous frequency band combination which maximizes the total information importance in the window with the screening frequency band number length based on a target function of a sliding window; and dividing the screened wavelet coefficient matrix into K subsets, training the intelligent frequency band screening collaborative network by adopting K-fold cross validation, and carrying out direct current bus capacitance parameter identification by utilizing the trained intelligent frequency band screening collaborative network. According to the invention, high-precision prediction can be carried out on the capacitance value of the capacitor and the equivalent series resistance at the same time under the condition of not additionally adding an additional circuit.

Description

Technical Field

[0001] This invention belongs to the field of power electronic equipment parameter detection technology, specifically a method for identifying DC bus capacitor parameters of DC-DC converters based on Discrete Wavelet Transform (DWT) and Intelligent Band Filtering Collaborative Network (IBS-CollNet). It is suitable for high-precision detection of DC bus output capacitance C and equivalent series resistance ESR in high-frequency power electronic systems. Background Technology

[0002] The DC bus capacitor is a core energy transfer component in a DC / DC conversion system, connecting the power converter to the output DC bus. It plays a crucial role in suppressing voltage fluctuations, maintaining bus voltage stability, and providing instantaneous power support. In high-frequency power electronic systems, the DC bus capacitor is highly susceptible to degradation or failure due to factors such as thermal stress and ripple current surges. When the output capacitance C decays or the equivalent series resistance (ESR) surges, the system's steady-state operation and dynamic response characteristics deteriorate significantly. During load transients, severe transient voltage drops are easily induced, weakening power quality, potentially causing control loop instability, and even irreversible physical damage to power devices. Therefore, achieving real-time, high-precision identification of the DC bus capacitance C and ESR without interfering with normal system operation is of significant engineering value for the lifecycle reliability of high-frequency power electronic systems.

[0003] Traditional modeling methods are sensitive to modulation parameters, have complex parameter tuning, and can only estimate single parameters, limiting their engineering application. Harmonic injection methods are only applicable to scenarios where harmonic injection is permitted, and are sensitive to control strategies, modulation methods, and load changes, resulting in insufficient robustness. Artificial intelligence (AI) methods have shown significant advantages in capacitor health monitoring, but they place higher demands on sampling channels and computational resources; moreover, they rely on manual experience for parameter tuning, which can easily lead to feature loss or redundancy, resulting in decreased model accuracy, reduced computational efficiency, and poor environmental adaptability. Summary of the Invention

[0004] The purpose of this invention is to propose a method for identifying DC bus capacitor parameters at the output of a DC-DC converter based on DWT and IBS-CollNet, in order to overcome the problems of existing DC bus capacitor parameter identification methods, such as strong reliance on human experience, the need for additional sampling channel components, difficulty in balancing identification accuracy and computational efficiency, and poor performance of multi-parameter synchronous identification.

[0005] The technical solution to achieve the objective of this invention is as follows: a method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet, comprising the following steps:

[0006] Step 1: Collect voltage ripple signal data across the DC bus capacitors under different aging conditions, and perform normalization preprocessing to eliminate the influence of dimensions, thus forming a normalized voltage ripple signal dataset.

[0007] Step 2: Perform discrete wavelet transform on the normalized voltage ripple signal to obtain a two-dimensional wavelet coefficient matrix containing information from multiple frequency bands;

[0008] Step 3: Construct a comprehensive evaluation mechanism for frequency band importance that integrates geometric and information entropy weighting, including: constructing an intelligent frequency band screening collaborative network; establishing a mapping relationship between the wavelet coefficient matrix and the predicted values ​​of capacitance and equivalent series resistance through the intelligent frequency band screening collaborative network; using the mean square error as the loss function; calculating the gradient of the loss function with respect to voltage ripple signal data using the chain rule; obtaining the normalized gradient importance of each frequency band with respect to capacitance and equivalent series resistance through batch sample gradient summation and normalization; and combining geometric weighting and information entropy weighting to solve for the comprehensive importance of each frequency band.

[0009] Step 4: Maximize the inter-class variance based on the frequency band comprehensive importance sequence to determine the optimal segmentation threshold and adaptively determine the number of key frequency bands to be selected; use the number of key frequency bands as the sliding window size to establish an objective function based on the importance of the sliding window; by moving the sliding scale window in the full frequency band space, find the continuous frequency band combination that maximizes the total information importance within the window, complete the screening of the wavelet coefficient matrix, and retain the time-frequency characteristics of the key frequency bands;

[0010] Step 5: Divide the selected wavelet coefficient matrix into K subsets, use K-fold cross-validation to train the intelligent band selection collaborative network, and use the trained intelligent band selection collaborative network to identify DC bus capacitor parameters.

[0011] Compared with the prior art, the significant advantages of this invention are:

[0012] 1. Requires only a single input and has low hardware cost: This invention uses the DC bus capacitor voltage ripple as the only identification input, without the need to simultaneously acquire current signals or add any additional sampling channels, detection elements or harmonic injection devices, which greatly reduces hardware cost and system complexity, avoids electromagnetic compatibility risks, and has low engineering implementation difficulty.

[0013] 2. Adaptive selection of key frequency bands to overcome reliance on human experience: A comprehensive evaluation mechanism for the importance of frequency bands, which integrates geometric and information entropy weighting, is designed. Combined with a threshold segmentation algorithm and a sliding window objective function, unsupervised and adaptive selection of key frequency bands is achieved. This completely overcomes the problems of poor environmental adaptability, feature loss or redundancy caused by manually preset frequency band thresholds in traditional methods; and effectively optimizes the allocation of computing resources.

[0014] 3. The dual-branch model architecture with attention mechanism has high recognition accuracy and strong robustness: The constructed IBS-CollNet model adopts a dual-branch parallel structure with small / large convolutional kernels, which can efficiently fuse local high-frequency detail features and global low-frequency time-frequency features. The embedded channel attention mechanism enhances the extraction of key frequency band information.

[0015] 4. Strong adaptability to operating conditions and wide range of applications: This invention does not rely on specific control strategies or modulation methods and has good adaptability to high-frequency power electronic systems with different loads and power ranges. It can realize real-time, high-precision monitoring of the entire life cycle of DC bus capacitors, providing an effective technical means for the reliability maintenance of power electronic systems.

[0016] 5. High computational efficiency and good real-time performance: By eliminating redundant frequency bands, the input dimension and computational load of the model are greatly reduced. The average training time of the IBS-CollNet model is only 49 seconds. Compared with traditional AI methods such as dual-input CNN and DCAE, the computational time and computing power requirements are greatly reduced, which can meet the computational needs of DC bus capacitor parameter identification.

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a flowchart of the DC bus capacitor parameter identification method based on DWT and IBS-CollNet proposed in this invention.

[0019] Figure 2 This is a schematic diagram of the main circuit topology of the bidirectional CLLC resonant converter according to an embodiment of the present invention.

[0020] Figure 3 This is an IBS-CollNet architecture diagram in the DC bus capacitor parameter identification method based on DWT and IBS-CollNet proposed in this invention.

[0021] Figure 4 This is a schematic diagram of the frequency band importance calculation process in the DC bus capacitor parameter identification method based on DWT and IBS-CollNet proposed in this invention.

[0022] Figure 5 This is a schematic diagram of the parameter identification prediction convergence result in the DC bus capacitor parameter identification method based on DWT and IBS-CollNet proposed in this invention.

[0023] Figure 6 This is a schematic diagram of the capacitance C parameter identification result in the DC bus capacitance parameter identification method based on DWT and IBS-CollNet proposed in this invention.

[0024] Figure 7 This is a schematic diagram of the identification results of the equivalent series resistance (ESR) parameter in the DC bus capacitance parameter identification method based on DWT and IBS-CollNet proposed in this invention.

[0025] Figure 8 This is a schematic diagram of the identification results of the capacitor degradation process by the frequency band screening mechanism in the DC bus capacitor parameter identification method based on DWT and IBS-CollNet proposed in this invention. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0028] The term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0029] A method for identifying DC bus capacitor parameters at the output of a DC-DC converter based on DWT and IBS-CollNet includes the following steps:

[0030] Step 1: Collect voltage ripple signal data across the DC bus capacitors under different aging conditions. The voltage ripple signal is then normalized to eliminate dimensionality effects, resulting in a normalized voltage ripple signal. Dataset;

[0031] DC-DC converters are a core topology in power electronic systems, converting one DC voltage level to another. They connect to other modules (DC-DC converters or DC-AC inverters) via a DC bus for power transfer. The DC bus capacitor, connected in parallel between the positive and negative terminals of the DC bus, plays a crucial role in suppressing voltage fluctuations, maintaining bus voltage stability, and providing instantaneous power support. In high-frequency power electronic systems, the DC bus capacitor is highly susceptible to degradation or failure due to factors such as thermal stress and ripple current surges. When the output capacitance C decreases or the equivalent series resistance (ESR) increases dramatically, the system's steady-state operation and dynamic response characteristics deteriorate significantly. To identify the aging degree of the DC bus capacitor at the DC-DC converter output, it is necessary to collect the DC bus capacitor voltage ripple signal under different aging conditions. After denoising and smoothing the signal, a normalization algorithm is used to eliminate the influence of dimensions, resulting in the preprocessed voltage ripple signal. This lays the foundation for subsequent DWT processing.

[0032] Step 2: Perform discrete wavelet transform on the normalized voltage ripple signal to obtain a two-dimensional wavelet coefficient matrix containing information from multiple frequency bands;

[0033] The preprocessed voltage ripple signal was subjected to discrete wavelet transform using the Dobessi 4 (db4) wavelet basis. This wavelet basis possesses the ability to accurately locate transient voltage changes and efficiently suppress high-frequency noise. It consists of a set of low-pass and high-pass filters of length 4. A certain number of decomposition levels J was set, and the voltage ripple signal was decomposed layer by layer using the low-pass and high-pass filters. After multiple decompositions, a set of wavelet coefficients was obtained, including the approximate coefficients of the highest decomposition level. and the detail coefficients obtained from each level of decomposition The length of each layer of coefficients is shortened due to the decomposition operation. To ensure that the length of the coefficients in each frequency band is consistent with the original signal for subsequent neural network processing, coefficient padding is performed on each group of wavelet coefficients. This is done by restoring the coefficient sequence to the length of the original voltage ripple signal through zero-value padding or interpolation padding. Finally, the one-dimensional voltage ripple signal is transformed into a two-dimensional wavelet coefficient matrix D, yielding the time-frequency characteristics of different frequency bands. The length of each frequency band is consistent with the length N of the original voltage ripple signal.

[0034] Step 3: Construct the IBS-CollNet model to calculate the comprehensive evaluation of the importance of frequency bands based on the fused geometry-information entropy weighting, and calculate the comprehensive importance F of the frequency bands. r A comprehensive evaluation mechanism for frequency band importance, integrating geometric and information entropy weighting, is constructed to achieve a quantitative evaluation of the characterization capability of each frequency band. The specific steps are as follows:

[0035] 3.1 Constructing the IBS-CollNet Model: This network adopts a dual-branch parallel structure with small convolutional kernels (3×5) and large convolutional kernels (5×7). The small convolutional kernels are used to extract local high-frequency detail features, while the large convolutional kernels are used to extract global low-frequency features. Channel attention mechanisms are embedded within each branch to enhance the feature extraction capability within the branch and suppress interference from invalid features. Both branches consist of convolutional layers, batch normalization layers, ReLU activation layers, attention mechanisms, and max-pooling layers stacked sequentially. The pooling layer ensures a uniform output scale between the two branches, facilitating subsequent feature concatenation. A random deactivation layer is added to each convolutional layer operation, randomly discarding some neurons to prevent overfitting. The features extracted by the parallel dual branches are flattened and concatenated to form a comprehensive feature vector that integrates local details and global information. The fused comprehensive feature vector is input to a fully connected layer, and after nonlinear mapping, the predicted values ​​of the output capacitance C and equivalent series resistance ESR are output.

[0036] 3.2: Establishing a neural network mapping relationship and calculating the importance of normalized gradients: A mapping relationship is established between the wavelet coefficient matrix D and the output capacitance value C and the equivalent series resistance ESR through a neural network. D is mapped to a two-dimensional output R². The mean square error (MSE) is used as the loss function 𝓛 to calculate the error between the predicted value and the true value. Using the chain rule, combined with the wavelet coefficient calculation formula and the voltage ripple expression, the gradient of the loss function with respect to the voltage ripple data is calculated to construct the mapping relationship between the actual capacitance parameters and the identification results. Gradient summation and normalization are performed on all batches of samples to obtain the normalized gradient importance of each frequency band with respect to C and ESR.

[0037] 3.3: Geometric Weighted G r Calculation: To ensure that the frequency band can simultaneously characterize both C and ESR parameters and avoid the dominance of a single parameter, a geometric weighted method is used to calculate G. r α represents the geometrically weighted importance of the frequency band with respect to parameter C, and β represents the geometrically weighted importance of the frequency band with respect to parameter ESR. The magnitude of these parameters characterizes the importance of the frequency band to the prediction. The importance of a frequency band for the two-parameter prediction is assigned through adaptive parameter tuning. Initially, the weights α and β are set to the standard geometric mean of 0.5. The optimal weight α is solved by minimizing the variance of the geometric weighting, ensuring a uniform distribution of importance among frequency bands. Then, the weight β is tuned to complete the prediction of G. r The calculation process involves adaptive calculation and optimization tuning of parameters, and ensures that the importance distribution between frequency bands is uniform, thereby optimizing the balance of frequency band evaluation.

[0038] 3.4: Information Entropy Weighted E r Calculation: To highlight the frequency bands that contribute significantly to parameter identification and are highly important, the information entropy weighted method is used to calculate E. rFirst, calculate the entropy H of the importance distribution of each frequency band to a single parameter. i H i A smaller value indicates that the parameter's prediction is highly dependent on a few frequency bands, and its importance is more concentrated; then, the maximum entropy H is determined based on the number of frequency bands. max Solve for the dispersion D i The degree of concentration of frequency band importance is evaluated; finally, based on the dispersion D... i The information entropy weighted importance weight w of parameter C c The information entropy weighted importance weight w of the parameter ESR ESR The frequency bands with high concentration of importance are selected through allocation. The calculation process involves adaptive calculation and optimization tuning of parameters, and the weighting of key frequency bands is strengthened.

[0039] 3.5: Overall Importance F r Solution: Combining geometric weighted G r Weighted E by information entropy r Solve for the overall importance F of each frequency band. r This enables a comprehensive evaluation of frequency band balance and criticality, and quantifies the contribution of each frequency band to C and ESR identification.

[0040] Step 4: Adaptive selection of key frequency bands based on threshold segmentation and sliding window, based on the comprehensive importance of frequency bands F. r A threshold segmentation algorithm combined with a sliding window objective function is used to achieve unsupervised and adaptive screening of key frequency bands. The specific steps are as follows:

[0041] 4.1: Threshold segmentation algorithm to determine the number of key frequency bands: Based on the threshold segmentation algorithm, according to the candidate segmentation threshold... In the frequency band importance sequence F r In the constructed set, frequency bands with importance greater than the threshold are assigned to the salient feature set, and frequency bands with importance less than the threshold are assigned to the redundant frequency band set. The probability weights of the salient feature set are then calculated. and the mean of overall importance Probability weights of redundant frequency band sets and the mean of overall importance When the inter-class variance is maximized, the difference between the two frequency bands is considered to be at its maximum. The threshold used to determine the overall importance of the two frequencies at this point is then considered the globally optimal threshold. The number of frequency bands in the salient feature frequency band set divided according to this threshold is the optimal number of frequency bands N. key While eliminating redundant components, it ensures the integrity of key physical characteristics over a wide power range.

[0042] 4.2: Sliding Window Selection of Continuous Key Frequency Bands: From the perspective of circuit physical characteristics, capacitor parameter shifts cause continuous shifts in the Laplace transform points of the converter output signal. The scaling characteristics of DWT map this physical evolution to the correlation between adjacent wavelet scales. From the perspective of neural network feature extraction requirements, discontinuous frequency bands cause breaks in the time-frequency image, destroying the spatial locality of features. Therefore, the number of key frequency bands N is selected. key The moving scale window length W is set, and the moving scale window moves across the entire frequency band space. The objective function of the sliding window is established with the goal of maximizing the comprehensive frequency band importance within the window. The wavelet coefficient matrix is ​​filtered, which can improve the model prediction accuracy while reducing the computation time and memory usage, thus enhancing the engineering practicality of the method.

[0043] Step 5: Input the filtered wavelet coefficient matrix, train it using the constructed IBS-CollNet model, and after K-fold model training and evaluation, achieve parameter identification and high-precision simultaneous identification of C and ESR. The specific steps are as follows:

[0044] The selected wavelet coefficient matrix is ​​divided into K subsets. During training, one subset is selected sequentially as the test set, and the remaining K-1 subsets are combined as the training set. A total of K fold model training and evaluation are performed. One training cycle involves inputting the training and test sets of the selected wavelet coefficient matrix into the IBS-CollNet constructed in step 3, and outputting the predicted values ​​of the corresponding output capacitance C and equivalent series resistance ESR. Mean squared error is used as the loss function to calculate the error between the predicted and true values. The network weights are continuously optimized to gradually reduce the loss function value until the model converges. The convergence result is shown below. Figure 5 As shown. The training and test sets of the wavelet coefficient matrix are repeatedly input into the neural network to output the predicted values ​​of the two parameters, completing one training epoch. This training process is repeated until the preset number of training epochs is reached, completing the training and evaluation of the one-fold model. After cross-validating the training and evaluation results of the K-fold model, the results are merged to obtain the recognition results for the entire dataset.

[0045] Example 1

[0046] This embodiment provides a method for identifying DC bus capacitance parameters based on DWT and IBS-CollNet, such as... Figure 1 As shown. This embodiment takes a typical high-frequency DC-DC converter, the bidirectional CLLC resonant converter, as an example, and its circuit topology is as follows. Figure 2As shown, the system includes, in sequence, the input terminal, the primary-side full-bridge inverter circuit, the resonant network and high-frequency isolation transformer circuit, the secondary-side full-bridge rectifier circuit, and the output terminal. The output terminal of the DC-DC converter (taking a CLLC resonant converter as an example) is connected to other modules in the power electronic system (such as a DC-AC inverter) via a DC bus for power transmission. The DC bus capacitor, connected in parallel between the positive and negative terminals of the DC bus, plays a crucial role in the DC-DC converter by suppressing voltage fluctuations, maintaining bus voltage stability, and providing instantaneous power support. In high-frequency DC-DC converters, this capacitor is more prone to degradation or failure due to factors such as thermal stress and ripple current surges compared to other passive components. Its performance stability directly determines the system's dynamic response capability and operational reliability. Studies have shown that when the output capacitance value C decays or the equivalent series resistance ESR surges, the system's steady-state operation and dynamic response characteristics will significantly deteriorate. Especially during load transients switching between half-load and full-load, capacitor parameter mismatch leads to a decrease in dynamic energy compensation capability, inducing severe transient voltage drops. This not only weakens the power quality of the system, but may also cause instability in the control loop, or even irreversible physical damage to power devices.

[0047] Using a 5kW CLLC resonant converter as an example of a DC-DC converter application, an experimental platform was built to verify the effectiveness of the method of this invention. The CLLC circuit parameters are shown in Table 1, and the input voltage... Output voltage Rated power 5kW, DC bus output capacitor value C identification range is 0.8 to 1 times the initial value of C, ESR is 1 to 2 times the initial value.

[0048] Table 1 CLLC circuit parameters used for parameter identification

[0049]

[0050] The method includes the following steps:

[0051] Step 1: Acquire and preprocess voltage ripple signal

[0052] On the DC bus of the CLLC converter, by replacing DC bus capacitors with different aging levels, the corresponding capacitor voltage ripple signals were collected. The collected voltage ripple signals were then sliced ​​into 500-byte segments, resulting in 199,836 training samples. These sliced ​​training samples were then denoised using a wavelet denoising algorithm to eliminate high-frequency noise interference; finally, a min-max normalization algorithm was used to refine the denoised signal. Normalization preprocessing is performed to map the signal values ​​to the [0,1] interval, eliminating the influence of dimensions. After the above operations, the preprocessed voltage ripple signal is obtained. .

[0053] Step 2: The preprocessed voltage ripple signal is subjected to discrete wavelet transform using the Dobessi 4 (db4) wavelet basis. A filter bank consisting of low-pass and high-pass filters decomposes the voltage ripple signal layer by layer. After passing through the wavelet basis, the voltage ripple signal is decomposed into wavelet coefficients layer by layer. The coefficients are then filled to restore the original voltage ripple signal length, and finally the one-dimensional voltage time-domain signal is transformed into a two-dimensional wavelet coefficient matrix D. The specific method is as follows:

[0054] First, the preprocessed voltage ripple signal is subjected to discrete wavelet transform using the Dobessi 4 (db4) wavelet basis. This wavelet basis possesses the ability to accurately locate transient voltage changes and efficiently suppress high-frequency noise. It consists of a set of wavelet bases of length L... db Low-pass filter with a value of 4 and high-pass filter Composition. The low-pass filter and the high-pass filter satisfy an orthogonal mirror relationship, that is:

[0055]

[0056] Where p represents the index of the filter coefficient, and its value range is... .

[0057] The normalized voltage ripple signal As an approximation coefficient for layer 0 Then, by continuously decomposing using low-pass and high-pass filters, the approximate coefficients of the j-th layer are obtained. and detail coefficient

[0058] , can be represented as

[0059]

[0060] Where m represents the displacement index of the wavelet coefficients, j is the index of the decomposition level, ranging from 1 to J; J is the decomposition level. Approximation coefficients and detail coefficient From the corresponding approximation coefficients After downsampling using low-pass and high-pass filters, the approximate coefficients are obtained by taking the index p of different filter coefficients, multiplying them using high-pass and low-pass filters respectively, and then summing the results. The length of the wavelet coefficients at each layer is shortened due to the downsampling decomposition operation. The original voltage ripple signal length N is restored by filling the coefficient sequence with zero values ​​or interpolation. Finally, the one-dimensional voltage ripple signal is transformed into a two-dimensional wavelet coefficient matrix D.

[0061]

[0062] In the formula, This represents the approximation coefficient of the j-th layer after filling. This represents the detail coefficient of the j-th layer after filling.

[0063] This matrix contains a set of approximate coefficients that characterize the low-frequency profile of the signal. and detail coefficients of multiple sets of signals representing high-frequency details at different scales In this embodiment, the length of the original voltage ripple signal is N=500. When the number of frequency bands after decomposition is selected as T=6, it corresponds to the decomposition to the 5th layer. The wavelet coefficient set includes the approximate coefficients of the 5th layer and the detail coefficients corresponding to each layer, resulting in the coefficients of 6 frequency bands (c5, d1-d5).

[0064] Step 3: Construct a comprehensive evaluation mechanism for frequency band importance that integrates geometric and information entropy weighting. This includes: constructing an intelligent frequency band selection collaborative network (IBS-CollNet) model for training; establishing a mapping relationship between the wavelet coefficient matrix and the capacitance value and equivalent series resistance through the intelligent frequency band selection collaborative network; using the mean square error as the loss function; calculating the gradient of the loss function with respect to voltage ripple data using the chain rule; and obtaining the normalized gradient importance of each frequency band for the capacitance value and equivalent series resistance by summing and normalizing the gradients of batch samples. Finally, combining geometric weighting and information entropy weighting, the comprehensive importance of each frequency band is solved. This step aims to automatically select the key frequency bands most important for the predicted output capacitance value C and equivalent series resistance ESR from the wavelet coefficient matrix generated by DWT. The specific method is as follows:

[0065] 3.1 Constructing the IBS-CollNet Model: This network adopts a dual-branch parallel structure with small convolutional kernels (3×5) and large convolutional kernels (5×7). The small convolutional kernels are used to extract local high-frequency detail features, while the large convolutional kernels are used to extract global low-frequency features. Channel attention mechanisms are embedded within each branch to enhance the feature extraction capability within the branch and suppress interference from invalid features. Both branches consist of convolutional layers, batch normalization layers, ReLU activation layers, attention mechanisms, and max-pooling layers stacked sequentially. The pooling layer ensures a uniform output scale between the two branches, facilitating subsequent feature concatenation. A random deactivation layer is added to each convolutional layer operation, randomly discarding some neurons to prevent overfitting. The features extracted by the parallel dual branches are flattened and concatenated to form a comprehensive feature vector that integrates local details and global information. The fused comprehensive feature vector is input to a fully connected layer, and after nonlinear mapping, the predicted values ​​of the output capacitance C and equivalent series resistance ESR are output.

[0066] 3.2 The wavelet coefficient matrix D after voltage ripple data processing in step 2 is used as the input data for the IBS-CollNet model. This input data contains 199,836 samples, each of which is a wavelet coefficient matrix of length N=500. During model training, the wavelet coefficient matrix samples are divided into different batches, with B samples in each batch. A neural network is used to establish the mapping relationship between the input wavelet coefficient matrix, the output capacitance value C, and the equivalent series resistance value ESR.

[0067]

[0068] In the formula This represents the wavelet coefficient matrix (length N * number of bandwidths T) as input to the neural network. This represents the two predicted values ​​output by the neural network. For the prediction parameter vector, This represents the true parameter vector. To measure the difference between the model's predicted values ​​and the actual label values, the mean squared error (MSE) is used as the loss function 𝓛. The mean squared error between the predicted and actual values ​​is calculated, and the network parameters are optimized using the backpropagation algorithm until the loss function converges, completing model training. The specific calculation method for the loss function 𝓛 is as follows:

[0069]

[0070] In the formula The predicted parameter vector includes the model-predicted capacitance value and the model-predicted equivalent series resistance value. The true parameter vector includes the true capacitance value and the true equivalent series resistance value of the DC bus capacitor. Using the chain rule, the gradient of the parameter identification loss function with respect to the voltage ripple data is calculated, and then the gradient importance of each frequency band is obtained. The specific calculation method is as follows:

[0071]

[0072] In the formula, the gradient can be expressed as the derivative of the loss gradient function φ with respect to the predicted value, the output predicted parameter vector of the IBS-CollNet model with respect to the two-dimensional input wavelet coefficient matrix, and the two-dimensional input wavelet coefficient matrix. Differentiation and wavelet coefficient matrix for normalized voltage ripple signal The product of derivatives.

[0073] By performing gradient summation and normalization on all batches of samples, the importance of the normalized gradient of each frequency band with respect to the two parameters can be determined. The specific calculation method is as follows:

[0074]

[0075] In the formula, Let r represent the gradient of frequency band r with respect to parameter i, B be the batch number set when inputting into the neural network, and N be the length of the original voltage ripple signal. After wavelet transform and padding, the length of the original signal is consistent with the length of the wavelet coefficients in the padded wavelet coefficient matrix D. By summing and normalizing the gradients of B batches with length N, the gradient importance of each frequency band with respect to the parameter to be predicted can be obtained. and .

[0076] To comprehensively evaluate the importance of a single frequency band, it is necessary to avoid simply adding its gradient importance with respect to the two parameters. Therefore, this invention designs a comprehensive frequency band importance calculation standard, which is divided into a geometrically weighted G that considers the balanced importance of the two frequency bands. r And considering the information entropy weighting E with high concentration of importance r The calculation method is as follows:

[0077]

[0078] In calculating geometrically weighted G r Weighted E by information entropy r The process involves several parameters to be tuned. To overcome the poor adaptability and subjectivity limitations caused by manual parameter tuning, this invention establishes an automatic parameter tuning mechanism that adaptively calculates and optimizes the parameters to be tuned based on the gradient importance of the parameters to be predicted.

[0079] In the geometric weighting method, it is required that the frequency band is equally important to both parameters, avoiding the dominance of a single parameter. Here, α is the geometrically weighted importance weight of the frequency band with respect to parameter C, and β is the geometrically weighted importance weight of the frequency band with respect to parameter ESR. Initially set to a standard geometric mean of 0.5, the weight parameters are optimized by minimizing the variance of the geometric weighting. The α with the smallest variance is taken as the optimal solution, and then the weight β is tuned to obtain a more uniform distribution of frequency band importance. The specific method for calculating the optimal weight is as follows:

[0080]

[0081] In the formula, α and β are two-parameter geometrically weighted importance parameters with tuning required for calculation, satisfying the following conditions: The relationship is determined by calculating the geometrically weighted mean. and variance With minimizing variance as the optimization objective, the α used in the calculation is the optimal solution, and then the weight β is obtained through tuning. The geometric weighted importance of the frequency bands for parameters C and ESR is initially set to 0.5. The optimal α is found to be 0.52 by minimizing the variance of the geometric weights, and β is tuned to 0.48 to ensure a uniform distribution of importance across the six frequency bands and avoid the dominance of a single parameter.

[0082] In the information entropy weighting method, frequency bands with highly concentrated importance are needed. c w is the weighted importance of the information entropy of the frequency band with respect to parameter C. ESR The information entropy weights of the frequency band with respect to the parameter ESR are weighted by importance. The required band tuning parameters need to be calculated, and the dispersion D needs to be calculated. i For weight w c and w ESR Distribute, D i The evaluation of whether the importance of frequency bands is concentrated, i.e. whether the parameters are determined by fewer frequency bands, is specifically calculated using the following method:

[0083]

[0084] In the formula, w c and w ESR To calculate the weighted importance of the entropy of the information to be tuned, we first calculate the normalized weights of the frequency band importance in terms of probability distribution. , i represents the parameter to be predicted, according to Calculate the entropy H of the importance distribution of each frequency band to a single parameter. i H i A smaller value indicates that the parameter's prediction is highly dependent on a few frequency bands, meaning its importance is concentrated, with the maximum entropy being H. max Then, through the entropy value H of the importance distribution i And the maximum entropy is H max Calculate the dispersion D i This includes the dispersion of parameter C, D. C The dispersion D of parameter ESR ESR Finally, based on the dispersion D i Weight w for the two information entropies c and w ESR The allocation is performed. The maximum entropy H is calculated. max =1.792, based on the weight allocation, highlighting the weight ratio of frequency band 3.

[0085] Calculate the geometric weighted G r Weighted E by information entropy r Then, the overall importance F of a single frequency band is calculated. r The overall calculation process is as follows: Figure 4 As shown. Combined with geometric weighting G r Weighted E by information entropy r The importance is comprehensively evaluated. The specific calculation method is as follows:

[0086]

[0087] In the formula, G is a geometrically weighted average. r Weighted E by information entropy rThe overall importance F of a single frequency band can then be obtained. r A comprehensive evaluation of the importance and criticality of frequency band balance was conducted, and the degree of importance of frequency band to the prediction results was weighed. The results are shown in Table 2.

[0088] Table 2 Frequency Band Importance Parameters

[0089]

[0090] Among them, band 3 has the highest overall importance, reaching a maximum score of 2.0, and makes the highest contribution to the identification of C and ESR; bands 2 and 4 are next, while bands 5 and 6 have relatively low importance and the lowest contribution, and are redundant bands.

[0091] Step 4: Adaptive screening of key frequency bands

[0092] Based on the threshold segmentation algorithm, the optimal segmentation threshold is determined by maximizing the inter-class variance of the comprehensive importance of frequency bands. The segmentation threshold can adaptively determine the number of key frequency bands to be selected.

[0093] The scale window length is determined based on the number of key frequency bands. The scale window is moved across the entire frequency band space to find the continuous frequency band combination that maximizes the overall importance of the frequency bands within the window, thus completing the screening of the wavelet coefficient matrix and preserving the time-frequency characteristics of the key frequency bands. This step addresses the limitations of traditional methods, such as poor environmental adaptability and feature loss due to manually preset screening thresholds. This invention introduces a frequency band screening strategy, including a threshold segmentation algorithm to determine the number of frequency bands and a sliding window objective function to achieve frequency band screening.

[0094] 4.1 Threshold Segmentation Algorithm: An unsupervised adaptive determination of the number of frequency bands is achieved based on the threshold segmentation algorithm. Its core logic lies in first determining the candidate segmentation thresholds. The system divides the frequency bands into a salient feature set and a redundant frequency band set. The overall importance of each frequency band is then placed into these sets according to different quantities and combinations. The inter-class variance of the two frequency band sets is calculated. When the inter-class variance is maximized, the difference between the two frequency bands is considered to be the greatest. The threshold used to divide the overall importance of the two frequency bands at this point is the globally optimal threshold. The number of frequency bands in the salient feature frequency band set divided according to this threshold is the optimal number of frequency bands. The specific calculation method is as follows:

[0095]

[0096] In the formula, As the candidate segmentation threshold, in the frequency band importance sequence F r In the constructed set, frequency bands with importance greater than the threshold are assigned to the salient feature set, and frequency bands with importance less than the threshold are assigned to the redundant frequency band set. The probability weights of the salient feature set are then calculated. and the mean of overall importance Probability weights of redundant frequency band sets and the mean of overall importance The optimal discrimination between target features and background interference can be evaluated through calculation; a higher optimal discrimination indicates a more significant distinction. Optimization is performed using the optimal discrimination as the objective function. After screening candidate segmentation thresholds, the candidate segmentation threshold with the highest discrimination is found, ultimately yielding the optimal segmentation threshold. The number of frequency bands in the salient feature frequency band set under this threshold is the optimal number of frequency bands N. key .

[0097] Compared to a fixed threshold, the threshold segmentation algorithm can dynamically adjust the number N of selected frequency bands according to the frequency band distribution characteristics under different operating conditions. key While eliminating redundant components with low identification contribution, it ensures the completeness of key physical characteristics over a wide power range.

[0098] The importance of frequency band integration F r Maximize the inter-class variance of the sequence to determine the optimal segmentation threshold =1.8, adaptively determining the number N of selected key frequency bands. key =3;

[0099] 4.2 Establishing the sliding window objective function: Given the optimal number of frequency bands N... key Furthermore, this invention considers the physical coherence between frequency bands. From a circuit perspective, when the capacitor C or the equivalent series resistance ESR shifts slightly, the Laplace transform point of the converter output signal in the complex plane will shift continuously. The scaling characteristics of the discrete wavelet transform can map this physical evolution to the correlation between adjacent wavelet scales. From a neural network perspective, the 2D convolutional kernel utilizes the frequency overlap characteristics of adjacent wavelet scales in the vertical dimension (frequency domain dimension) to extract scale gradient features by capturing the evolution trend of the scale. If the frequency band selection is discontinuous, it will cause the time-frequency image perceived by the neural network to be fragmented, destroying the spatial locality of the features.

[0100] The number of key frequency bands N key Setting the scale window length as W, the scale window moves across the entire frequency band. A sliding window objective function is established with the optimization objective of maximizing the overall frequency band importance within the window. The wavelet coefficient matrix is ​​then filtered. This approach improves model prediction accuracy while reducing computation time and memory usage, enhancing the method's engineering practicality. The specific calculation method for the sliding window-based objective function is as follows:

[0101]

[0102] With a window size W=3, the window is moved within 6 frequency bands, and the total information importance within each window is calculated. The results show that when the window covers frequency bands 2-4, At this point, the total information is of the greatest importance, so frequency bands 2-4 are selected as continuous key frequency bands to complete the screening of wavelet coefficient matrix D.

[0103] Step 5: Divide the filtered wavelet coefficient matrix into K subsets, and train the intelligent band filtering collaborative network using K-fold cross-validation. Then, use the trained intelligent band filtering collaborative network to identify DC bus capacitor parameters. Specifically:

[0104] The selected wavelet coefficient matrix is ​​divided into K subsets. During training, one subset is selected sequentially as the test set, and the remaining K-1 subsets are combined as the training set. A total of K fold model training and evaluation are performed; in this embodiment, K is set to 5. One training process involves inputting the training and test sets of the selected band 2-4 wavelet coefficient matrix into the IBS-CollNet constructed in step 3, outputting the predicted values ​​of the corresponding output capacitance C and equivalent series resistance ESR. Mean squared error is used as the loss function to calculate the error between the predicted and true values. The Adam optimizer is used with a learning rate of 0.001 to continuously optimize the network weights, gradually reducing the loss function value until the model converges. The convergence result is shown below. Figure 5 As shown. The training and test sets of the wavelet coefficient matrix are repeatedly input into the neural network to output the predicted values ​​of the two parameters, completing one training epoch. This training process is repeated until the preset number of training epochs (150 epochs) is reached, completing the training and evaluation of the one-fold model. After cross-validation of the K-fold model training and evaluation results, the results are merged to obtain the identification results for the entire dataset. The identification results for C and ESR are shown below. Figure 6 and Figure 7 As shown, Figure 6 and Figure 7 In the diagram, the red dashed line represents the ideal fitting line where the predicted value equals the true value, and the blue scatter points represent the neural network's predictions for several samples. The distance between the blue scatter points and the red dashed line represents the deviation between the predicted and true values. In the figure, all the blue scatter points are near the red dashed line, and the data points are highly linearly distributed along the ideal fitting line, indicating that this method can accurately predict the parameters.

[0105] The results show that the method of this invention achieves high-precision simultaneous identification of two key parameters, C and ESR, during capacitor degradation without introducing additional components. The prediction accuracy of C reaches 96.74%, and the prediction accuracy of ESR is 97.71%, demonstrating that the model's predicted values ​​have a strong interpretability of the actual capacitor parameters. Simultaneously, this method significantly reduces the computational burden, achieving a better balance between estimation accuracy and training efficiency.

[0106] The present invention is specifically verified and compared, including:

[0107] Comparison of frequency band selection methods: When the number of DWT decomposition frequency bands was set to 8, the effectiveness of the frequency band selection mechanism in identifying the capacitor degradation process was further tested. The results are illustrated in the diagram below. Figure 8 As shown in the figure, the network trained in the proposed method predicts the capacitance and ESR values ​​of each input tensor. Therefore, the average of the five predictions is used as the final result to reduce the impact of each estimation error. The performance on the test set is compared with different numbers of frequency bands and frequency band selection strategies, and the results are shown in Table 3.

[0108] Table 3 Performance Comparison of Different Number of Frequency Bands and Frequency Band Selection Strategies

[0109]

[0110] Table 3 compares the impact of continuous and discontinuous frequency band selection strategies on model training efficiency and the identification performance of output capacitance value C and equivalent series resistance ESR for DC bus capacitance parameter identification tasks with four frequency band counts of 8, 6, 4, and 3. The results are presented in terms of training time and the coefficient of determination C_R. 2 With ESR_R 2 The relative root mean square error (RRMSE) and ESR_RRMSE were used as the core evaluation metrics to comprehensively verify the superiority of the continuous frequency band selection strategy and the optimization effect of redundant frequency band removal on model performance. In terms of training time, as the number of frequency bands decreased, the training time of both continuous and discontinuous strategies showed a significant downward trend. The training time for the 8-band continuous strategy was 126.3s, which was reduced to only 49.0s for the 3-band continuous strategy, a reduction of over 60%. The discontinuous strategy also showed the same downward trend, proving that removing redundant frequency bands can effectively reduce the computational load of the model, reduce the demand for computing resources, and improve training efficiency. In terms of recognition performance, the continuous frequency band selection strategy outperformed the discontinuous strategy with the same number of frequency bands across all frequency band counts. Under the 8-band continuous strategy, the RRMSE was significantly lower than that of the 8-band continuous strategy. 2 ESR_R 2 The values ​​are 0.9482 and 0.9506 respectively, and C_RRMSE and ESR_RRMSE are 0.0580 and 0.0548, while the C_R of the discontinuous strategy in the same frequency band is... 2 ESR_R 2The C_RRMSE and ESR_RRMSE values ​​decreased to 0.8550 and 0.8902, respectively, while C_RRMSE and ESR_RRMSE increased to 0.0986 and 0.0829, respectively. All core metrics showed significant deterioration, a trend consistent across frequency bands 6, 4, and 3. This demonstrates that frequency band continuity is crucial for ensuring model recognition accuracy. Discontinuous frequency band selection disrupts the spatial locality and physical coherence of time-frequency features, preventing the neural network from effectively capturing relevant features of capacitance parameters. Conversely, under the continuous frequency band strategy, as the number of frequency bands gradually decreased from 8 to 3, the model's recognition performance showed a continuous optimization trend, with C_R... 2 The value was increased from 0.9482 to 0.9610, ESR_R 2 The improvement from 0.9506 to 0.9676, the decrease in C_RRMSE from 0.0580 to 0.0506, and the decrease in ESR_RRMSE from 0.0548 to 0.0450 demonstrate that the frequency band selection mechanism proposed in this invention can effectively eliminate redundant frequency bands with low contribution to parameter identification, retain key time-frequency features, and reduce interference from invalid features, thereby improving the model's identification accuracy. In summary, the 3-band continuous selection strategy is the optimal solution. It not only compresses the training time to 49.0s, maximizing computational efficiency, but also achieves the highest identification accuracy. All performance indicators are optimal, verifying the scientific validity and effectiveness of the sliding window-based continuous key frequency band selection method in this invention. It achieves reasonable optimization of computational resources while ensuring model identification performance.

[0111] Comparison of parameter identification performance of different models: Most traditional signal processing and capacitance estimation methods have certain limitations in model generalization ability and simultaneous detection of two parameters. In contrast, the method of this invention achieves high-precision simultaneous identification of two key parameters, C and ESR, during capacitance degradation without introducing additional components. The prediction accuracy of C reaches 96.74%, and the prediction accuracy of ESR is 97.71%, proving that the model's predicted values ​​have a strong interpretability of the actual capacitance parameters. At the same time, this method significantly reduces the computational burden, achieving a better balance between estimation accuracy and training efficiency.

[0112] This paper comprehensively compares the performance differences of the proposed method with existing methods across six dimensions: identification components, dual-parameter identification capability, whether additional components are required, computational burden, adaptability to topology, and other application limitations. The performance of each method in addressing core technical challenges is quantified using "S" (solved) and "U" (unsolved). The system verifies the comprehensive advantages of the proposed method in DC bus capacitor parameter identification. A performance comparison with existing methods is summarized in Table 4.

[0113] Table 4 Performance Comparison of Different Models and Number of Frequency Bands

[0114] In terms of modeling methods, existing technologies can only identify a single parameter (C or ESR) and cannot perform simultaneous detection of both C and ESR parameters. Furthermore, some methods require additional components, and others suffer from excessive computational burden. The adaptable topologies are also limited to specific structures such as Buck and PFC, lacking universal adaptability and exhibiting significant technical limitations.

[0115] Regarding injection methods, although a few methods can achieve dual parameter identification of C and ESR without the need for additional components, these methods generally suffer from heavy computational burden and have strict application limitations, such as only being applicable to PWM rectifiers with front-end rectifiers and requiring the use of DC reactors. They cannot be widely applied in conventional power electronic topologies, and harmonic injection may bring electromagnetic compatibility risks, further limiting their engineering practicality.

[0116] In terms of traditional artificial intelligence methods, although existing technologies have solved the problem of simultaneous identification of two parameters without the need for additional components and have expanded the adaptability to topologies, they all share the common problem of excessive computational burden. Some methods also require the simultaneous acquisition of voltage and current signals, which increases the hardware cost of the sampling channel and the computational cost of data processing, resulting in certain resource consumption issues in engineering applications.

[0117] The proposed DC bus capacitor parameter identification method based on DWT and IBS-CollNet achieves "S" (Solution) across all core evaluation dimensions. It not only enables high-precision simultaneous identification of both C and ESR parameters without requiring additional sampling elements, detection devices, or harmonic injection equipment, significantly reducing hardware costs and system complexity, but also effectively addresses the computational burden of existing methods. Through adaptive key frequency band selection and structural optimization of the IBS-CollNet model, it significantly reduces the computational load and training time of the model. Furthermore, it is compatible with mainstream topologies of high-frequency power electronic systems such as CLLC resonant converters, with no additional application limitations, demonstrating significant engineering practicality. In summary, this method overcomes the technical bottlenecks of existing model methods, injection methods, and traditional artificial intelligence methods in terms of parameter identification capability, hardware requirements, computational efficiency, and topology adaptability. It is a DC bus capacitor parameter identification method more suitable for practical engineering applications, exhibiting significant innovation and practicality compared to existing technologies.

[0118] Performance Comparison of Parameter Identification with Different Numbers of Artificial Intelligence and Frequency Bands: To further verify the performance advantages of the method of this invention compared with artificial intelligence algorithms, we used various artificial intelligence algorithms to compare the performance of C and ESR identification tasks, including CNN models and DCAE models with different numbers of frequency bands. All algorithms were executed offline on a computer equipped with an AMD RX 7600 to ensure that the comparison was not affected by other factors. Relative root mean square error (RRMSE), coefficient of determination (R²), and accuracy were used as evaluation indicators to comprehensively evaluate the parameter identification effect. The comparison results of each model are shown in Table 5.

[0119] Table 5 Performance Comparison of Different Models and Frequency Band Quantities

[0120]

[0121] Table 5 compares the DC bus capacitance parameter identification performance of traditional CNN models in 6-band and 3-band, DCAE, and the IBS-CollNet model proposed in this invention in 6-band and 3-band. It fully verifies the structural superiority of the IBS-CollNet model and the positive optimization effect of key band selection on model performance. At the same time, it highlights the performance defects of traditional CNN models after band reduction and the computational efficiency shortcomings of the DCAE model.

[0122] From the perspective of training time, the three types of models exhibit distinct efficiency differences. Reducing the number of frequency bands only positively improves training efficiency for CNN and IBS-CollNet models. The DCAE model, due to the complexity of its network structure, suffers from inherent limitations due to excessive computational burden and lacks explicit frequency band adaptation optimization design. Its overall training time is significantly higher than that of CNN and IBS-CollNet models under the same task, becoming a core limitation for its engineering applications. Conversely, for CNN and IBS-CollNet models, reducing the number of frequency bands improves training efficiency for all models. Furthermore, the training time of the IBS-CollNet model with the same number of frequency bands is significantly lower than that of traditional CNN models: the traditional CNN model's training time for 6 frequency bands is 123.96s, while for 3 frequency bands it drops to 73.7s, a reduction of approximately 40.6%; the IBS-CollNet model's training time for 6 frequency bands is only 89.5s, a 27.8% reduction compared to the traditional CNN with the same frequency bands, and for 3 frequency bands it further drops to 49.0s, significantly lower than the traditional CNN with the same frequency bands. The training time was reduced by 33.7% and by 45.2% compared to its own 6-band training time, which fully demonstrates that the dual-scale convolutional parallel structure of the IBS-CollNet model combined with the channel attention mechanism has optimized the computational efficiency in terms of model structure, while the key frequency band selection further eliminates redundant computation. The two work together to achieve a significant improvement in training efficiency.

[0123] From the perspective of recognition performance, the three types of models show significant differences. While the DCAE model can achieve dual-parameter recognition of C and ESR with high accuracy, overcoming the limitations of traditional single-parameter recognition methods, its complex network structure sacrifices computational efficiency for high accuracy, making it difficult to balance recognition performance with training time. Traditional CNN models, while possessing some recognition capability across six frequency bands, also exhibit high accuracy in C and ESR recognition. 2 ESR_R 2 The C and ESR RMSE values ​​reached 0.9472 and 0.9513 respectively, with C_RRMSE and ESR_RRMSE values ​​of 0.0576 and 0.0545. The identification accuracies for C and ESR also reached 91.14% and 93.00% respectively. However, when the number of frequency bands was reduced from 6 to 3, its identification performance experienced a precipitous drop, with C_R... 2 ESR_R 2 The C_RRMSE and ESR_RRMSE values ​​dropped to 0.4028 and 0.3766, respectively, while the C_RRMSE and ESR_RRMSE values ​​rose to 0.1936 and 0.1948, respectively. The recognition accuracy plummeted to 34.24% and 35.6%, respectively, with all core metrics showing significant deterioration. This demonstrates that traditional CNN models lack the ability to effectively mine and utilize frequency band features. The reduction in the number of frequency bands directly leads to the loss of key features, making effective parameter identification impossible. In contrast, the IBS-CollNet model proposed in this invention maintains excellent recognition performance throughout the frequency band reduction process and achieves continuous performance optimization: the C_RRMSE of the IBS-CollNet model under 6 frequency bands... 2 ESR_R 2 The C_RRMSE and ESR_RRMSE have reached 0.9550 and 0.9554 respectively, while the C_RRMSE and ESR_RRMSE have decreased to 0.0543 and 0.0527. The recognition accuracy of C and ESR is 91.05% and 93.33% respectively, and all indicators are slightly better than the traditional 6-band CNN model. When the number of bands is reduced to 3, its recognition performance is further improved, and the C_R... 2 ESR_R 2 The C_RRMSE and ESR_RRMSE values ​​increased to 0.9610 and 0.9676 respectively, while the C_RRMSE and ESR_RRMSE values ​​decreased to 0.0506 and 0.0450 respectively. The C_RRMSE recognition accuracy improved to 94.30%, and the ESR_RRMSE recognition accuracy improved significantly to 97.19%. All core indicators reached their optimal levels.

[0124] This result fully demonstrates that the dual-scale convolutional parallel structure of the IBS-CollNet model can efficiently extract local high-frequency detail features and global low-frequency features. The embedded channel attention mechanism can adaptively enhance the feature information of key frequency bands. Combined with the geometric-information entropy weighted frequency band filtering mechanism proposed in this invention, it can accurately retain time-frequency features that make core contributions to C and ESR identification, effectively eliminate invalid interference from redundant frequency bands, fundamentally solve the problem of poor frequency band adaptability of traditional CNN models, and overcome the shortcoming of the DCAE model that "high accuracy and low efficiency cannot be achieved simultaneously". In summary, the IBS-CollNet model with 3 frequency bands is the optimal solution for online parameter identification of DC bus capacitors. While compressing the training time to 49.0s and maximizing computational efficiency, it achieves identification accuracy that is far superior to traditional CNN models, comparable to the DCAE model, and more efficient, thus realizing the dual goals of "improving computational efficiency" and "optimizing identification accuracy". Compared to the frequency band adaptation defects of traditional CNN models and the computational efficiency shortcomings of DCAE models, the IBS-CollNet model, which combines a key frequency band screening mechanism, has achieved a qualitative improvement in recognition performance, computational efficiency, and frequency band adaptation, fully verifying the innovation, rationality, and engineering practicality of the model design of this invention.

[0125] Compared with existing technologies, the DC bus capacitor online parameter identification method based on DWT and IBS-CollNet adopted in this embodiment uses only voltage ripple as a single input to achieve high-precision synchronous identification of C and ESR, overcomes the dependence on manual experience, and greatly improves the computational efficiency. It has good application results in CLLC resonant converters and can be extended to the DC bus capacitor parameter detection of other high-frequency power electronic systems.

[0126] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. For example, the strategy proposed according to the present invention can be combined with other optimization algorithms to adaptively adjust the damping coefficient. All variations that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should be considered within the scope of protection of this invention.

Claims

1. A method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet, characterized in that, Includes the following steps: Step 1: Collect voltage ripple signal data across the DC bus capacitors under different aging conditions, and perform normalization preprocessing to eliminate the influence of dimensions, thus forming a normalized voltage ripple signal dataset. Step 2: Perform discrete wavelet transform on the normalized voltage ripple signal to obtain a two-dimensional wavelet coefficient matrix containing information from multiple frequency bands; Step 3: Construct a comprehensive evaluation mechanism for frequency band importance that integrates geometric and information entropy weighting, including: constructing an intelligent frequency band screening collaborative network; establishing a mapping relationship between the wavelet coefficient matrix and the predicted values ​​of capacitance and equivalent series resistance through the intelligent frequency band screening collaborative network; using the mean square error as the loss function; calculating the gradient of the loss function with respect to voltage ripple signal data using the chain rule; obtaining the normalized gradient importance of each frequency band with respect to capacitance and equivalent series resistance through batch sample gradient summation and normalization; and combining geometric weighting and information entropy weighting to solve for the comprehensive importance of each frequency band. Step 4: Maximize the inter-class variance based on the frequency band comprehensive importance sequence to determine the optimal segmentation threshold and adaptively determine the number of key frequency bands to be selected; use the number of key frequency bands as the sliding window size to establish an objective function based on the importance of the sliding window; by moving the sliding scale window in the full frequency band space, find the continuous frequency band combination that maximizes the total information importance within the window, complete the screening of the wavelet coefficient matrix, and retain the time-frequency characteristics of the key frequency bands; Step 5: Divide the selected wavelet coefficient matrix into K subsets, use K-fold cross-validation to train the intelligent band selection collaborative network, and use the trained intelligent band selection collaborative network to identify DC bus capacitor parameters.

2. The method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet according to claim 1, characterized in that, Discrete wavelet transform is performed on the preprocessed voltage ripple signal using the Dobessi 4 wavelet basis, which includes a low-pass filter. and high-pass filter ; The normalized voltage ripple signal As an approximation coefficient for layer 0 Then, by continuously decomposing using low-pass and high-pass filters, the approximate coefficients of the j-th layer are obtained. and detail coefficient Specifically: In the formula, p is the index of the filter coefficient. The total length of the low-pass and high-pass filters is denoted by m, which represents the displacement index of the wavelet coefficients. j is the index of the decomposition level, ranging from 1 to J; J is the decomposition level. The approximation coefficient and detail coefficient sequences are restored to the original voltage ripple signal length N by zero-filling or interpolation filling; The two-dimensional wavelet coefficient matrix D is constructed using the detail coefficients and the approximation coefficients of the last decomposition layer, as follows: In the formula, This represents the approximation coefficient of the j-th layer after filling. This represents the detail coefficient of the j-th layer after filling.

3. The method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet according to claim 2, characterized in that, The low-pass filter and the high-pass filter satisfy an orthogonal mirror relationship, that is: 。 4. The method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet according to claim 1, characterized in that, The intelligent frequency band selection collaborative network takes a two-dimensional wavelet coefficient matrix as input. The network includes a small convolutional kernel branch and a large convolutional kernel branch. Both branches are composed of convolutional layers, batch normalization layers, ReLU activation layers, attention mechanisms, and max pooling layers stacked sequentially. The small convolutional kernel branch is used to extract local high-frequency detail features, while the large convolutional kernel branch is used to extract global low-frequency features. The features extracted by the small and large convolutional kernel branches are flattened and fused to form a comprehensive feature vector that integrates local details and global information. The fused comprehensive feature vector is input to a fully connected layer, and after nonlinear mapping, the predicted values ​​of the output capacitance C and equivalent series resistance ESR are output.

5. The method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet according to claim 1, characterized in that, The specific formula for calculating the loss function 𝓛 is as follows: In the formula, This is the actual value of the output capacitance or equivalent series resistance. The output capacitance or equivalent series resistance is the predicted value, where C is the output capacitance and ESR is the equivalent series resistance. The gradient of the loss function for parameter identification with respect to voltage ripple data is calculated using the chain rule, and then the gradient importance of each frequency band is obtained. The specific calculation formula is as follows: In the formula, D is the two-dimensional wavelet coefficient matrix. Explanation of parameters for the preprocessed voltage ripple signal. Gradient summation and normalization were performed on all batches of samples to obtain the normalized gradient importance of each frequency band to capacitance and equivalent series resistance. The specific calculation formula is as follows: In the formula, The gradient of the r-band with respect to the i-parameter is represented by B, the batch number set when inputting into the neural network is B, and the length of the original voltage ripple signal is N. After wavelet transform and padding, the lengths of the approximation coefficients and detail coefficients in the wavelet coefficient matrix D are consistent with the length of the original voltage ripple signal. Combined with geometric weighting Weighted by information entropy The formula for calculating the overall importance of each frequency band is as follows: In the formula, The importance of the information entropy of the frequency band with respect to the output capacitance value C is weighted. α is the information entropy weighted importance weight of the frequency band with respect to the parameter ESR, β is the geometric weighted importance weight of the frequency band with respect to the output capacitance value C, and β is the geometric weighted importance weight of the frequency band with respect to the equivalent series resistance value ESR.

6. The method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet according to claim 5, characterized in that, The weights α and β are determined as follows: In the formula, The geometrically weighted mean. Let r be the variance, r be the frequency band index, and the value range is 1-T. T represents the total number of frequency bands, and T=J+1, where J is the number of decomposition layers of the discrete wavelet transform. , The method for determining it is as follows: In the formula, H is the normalized weight of frequency band importance in the sense of probability distribution. i H represents the entropy value of the importance distribution of each frequency band to a single parameter. max D represents the maximum entropy value of the importance distribution of each frequency band to a single parameter. C D represents the dispersion of the output capacitance value C. ESR This represents the dispersion of the equivalent series resistance (ESR).

7. The method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet according to claim 1, characterized in that, The specific formula for determining the optimal segmentation threshold based on maximizing the inter-class variance according to the frequency band comprehensive importance sequence is as follows: In the formula, The candidate segmentation threshold, The probability weights for the salient feature set. The probability weights for the redundant frequency band set. The mean of the overall importance of the salient feature set. This represents the average overall importance of the redundant frequency band set; The number of frequency bands in the set of salient feature frequency bands divided according to the optimal segmentation threshold is the optimal number of frequency bands.

8. The method for identifying DC bus capacitor parameters at the output terminal of a DC-DC converter based on DWT and IBS-CollNet according to claim 1, characterized in that, Based on the number of key frequency bands N t Determine the scale window length W, move the scale window across the entire frequency band, and establish a sliding window objective function with the optimization objective of maximizing the comprehensive frequency band importance within the window. Filter the wavelet coefficient matrix. The specific formula for calculating the sliding window objective function is as follows: In the formula, W represents the window size, indicating the overall importance of each frequency band.

Images