A method and system for degradation fault output voltage ripple optimization for a dc-dc converter

Abstract

The application discloses a degradation fault output voltage ripple optimization method and system for a DC-DC converter, relates to data processing processes such as undersampling, feature extraction and feature selection, a fault diagnosis method based on an extreme learning machine and a reliability optimization method based on DC-DC converter model analysis, belongs to the field of power electronic system fault diagnosis, and comprises the following steps: undersampling technology for the output voltage of a DC-DC converter; time domain feature extraction technology for undersampling data of the output voltage of the DC-DC converter; recursive feature elimination technology based on a support vector machine for selecting the time domain features of the extracted undersampling data of the output voltage of the DC-DC converter; a support vector machine degradation fault diagnosis model based on a Gaussian kernel and a Bayesian optimization for learning multi-class degradation fault data and performing fault classification; and an analysis method based on a physical model of a power conversion device for optimizing the output voltage ripple of the power conversion device.

Description

Technical Field

[0001] This invention belongs to the field of power electronic system fault diagnosis, and more specifically, relates to a method and system for optimizing the output voltage ripple of a DC-DC converter in the event of a degradation fault. Background Technology

[0002] Advances in power electronics technology and the gradual popularization of third-generation semiconductor devices have spurred the development of DC-DC converters. Against the backdrop of a strong push for green development, the demand for DC-DC converters in energy storage systems, aerospace, DC microgrids, renewable energy, and new energy vehicles is also increasing. At the same time, this places higher demands on the performance indicators of DC-DC converters.

[0003] Output capacitors and filter inductors in DC-DC converters are more susceptible to degradation due to high temperatures, humidity, vibration, and other environmental stresses. When degradation reaches a certain level, the converter is considered to have experienced a soft fault. Degradation of electrolytic capacitors is mainly manifested as a decrease in capacitance (C) or an increase in equivalent series resistance (ESR). When C decreases by 20% or ESR increases by 180-200%, the capacitor is considered to have failed. Inductor failure is characterized by a 20% decrease in inductance (L) or a 100% increase in ESR. To prevent system failures from endangering personal and property safety, it is necessary to assess the degradation status in advance for timely maintenance. Existing DC-DC converter fault diagnosis methods are mainly divided into several categories, including model-based methods, signal-based methods, and knowledge-based methods, the latter further defined as data-driven methods.

[0004] Model-based methods analyze the fault mechanisms of DC-DC converters and establish mathematical models for soft fault diagnosis. However, these methods are only applicable to specific converter structures, and the diagnostic accuracy depends on the appropriateness of the established model. Signal-based methods typically diagnose soft faults in DC-DC converters by comparing the signals of the converter's current state with those of its normal state. However, this method requires setting reasonable fault thresholds and relies on empirical judgment and the application scenario. Data-driven methods, including statistical and machine learning methods, do not require knowledge of the converter's structure and fault mechanisms. These methods can achieve good performance on existing datasets, but solutions have not yet been proposed for industrial low-computing-power scenarios. Furthermore, how to guide converter control and optimize its reliability based on diagnostic results is a real challenge. Additionally, in practical applications, it is difficult to obtain high-sampling-frequency output voltage signals from the converter, often affecting diagnostic accuracy and failing to meet the data requirements of the model. Low sampling frequency is also a problem that cannot be ignored in applications. At the same time, low-frequency sampling can significantly reduce system costs. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for optimizing the output voltage ripple of a DC-DC converter in the event of degradation faults. This method and system are designed to effectively diagnose soft faults in DC-DC converters in industrial scenarios with low computing power and low sampling frequency, and to optimize the performance of the DC-DC converter in a timely manner through reliability optimization strategies.

[0006] This invention provides a method for optimizing the degradation fault output voltage ripple of a DC-DC converter, comprising the following steps:

[0007] S1. Undersample the output voltage of the DC-DC converter to obtain the output voltage ripple;

[0008] S2. Extract the time-domain features of the output voltage ripple to form a feature dataset;

[0009] S3. Use recursive feature elimination based on support vector machine to select the optimal number of optimal features and form the optimal feature dataset;

[0010] S4. The fault diagnosis model is trained using the optimal feature dataset. The fault diagnosis model is based on an extreme learning machine. The hyperparameters of the fault diagnosis model are optimized through multiple iterative training cycles to obtain the optimal combination of hyperparameters, thereby obtaining the trained fault diagnosis model.

[0011] S5. Input the real-time output voltage of the DC-DC converter to the trained fault diagnosis model to perform real-time diagnosis of the DC-DC converter;

[0012] S6. Based on the real-time diagnostic results of the DC-DC converter, infer the degree of degradation of the capacitors and inductors of the DC-DC converter, adjust the switching frequency, and reduce the output voltage ripple.

[0013] Furthermore, the time-domain features include maximum value, minimum value, peak value, peak-to-peak value, average value, absolute average value, square root amplitude, variance, standard deviation, root mean square, kurtosis, skewness, shape factor, peak factor, impulse factor, margin factor, and gap coefficient.

[0014] Furthermore, the recursive feature elimination based on support vector machines includes: establishing a sample training model, scoring and ranking each temporal feature, deleting the temporal feature with the lowest feature score, and then retraining the model using the remaining temporal features for the next iteration until the required number of optimal features are selected; wherein the ranking criterion score of temporal feature i is defined as:

[0015]

[0016] Where i = 1, 2, ..., n, ω i denoted as the weight of time-domain feature i.

[0017] Furthermore, the frequency of undersampling is set to Where η is any positive integer, and f is the switching frequency of the DC-DC converter.

[0018] Furthermore, the training steps of the fault diagnosis model are as follows: An ELM network is selected as the basic network model for fault diagnosis; the network structure of the ELM is first randomly initialized, and the ELM network parameters are trained using a subset of the optimal feature datasets as the training set. During training, the model gradually adjusts the network parameters using the least squares method to better fit the training data; then, the remaining optimal feature datasets are used as the validation set to verify the diagnostic accuracy under this network structure; based on the validation results of the validation set, the ELM network hyperparameters that best fit the fault diagnosis task dataset are found through continuous iterative optimization using the Bayesian optimization algorithm, i.e., the optimal network structure is found; finally, based on the optimal network structure, the training set is used for training to obtain the trained fault diagnosis model.

[0019] Furthermore, the Bayesian optimization algorithm includes:

[0020] Step 1: Initialize the Gaussian process with prior knowledge based on any available prior knowledge;

[0021] Step 2: Use the acquisition function to determine the next sampling point based on the current Gaussian process posterior distribution;

[0022] Step 3: Evaluate the objective function at the selected point and update the Gaussian process to obtain a new posterior distribution;

[0023] Step 4: Repeat steps 2-3, using the updated posterior as the new prior for the Gaussian process.

[0024] Furthermore, the peak-to-peak value of the output voltage ripple of the DC-DC converter is expressed as:

[0025]

[0026] Where f is the switching frequency, L is the filter inductance, C is the output capacitor, and V is the voltage. in V is the input voltage. o This is the output voltage.

[0027] Given the isolated application problem between existing fault diagnosis and reliability optimization methods, which prevents them from achieving synergy and optimization, the purpose of this invention is to:

[0028] 1. Develop a new fault diagnosis method that can accurately identify different fault conditions in DC-DC converters.

[0029] 2. Establish a correlation model between diagnostic results and control parameters to achieve an efficient combination of fault detection and reliability optimization.

[0030] 3. Based on the fault situation, dynamically adjust the operating parameters of the DC-DC converter to reduce the impact of the fault on system performance.

[0031] 4. To achieve intelligent fault prediction, diagnosis and control of DC-DC converters, thereby improving system reliability and availability.

[0032] 5. To provide a new technical means for condition monitoring and fault management of DC-DC converters.

[0033] The present invention also provides a degradation fault output voltage ripple optimization system for a DC-DC converter, comprising: a computer-readable storage medium and a processor;

[0034] The computer-readable storage medium is used to store executable instructions;

[0035] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the above-described method for optimizing the degradation fault output voltage ripple of a DC-DC converter.

[0036] Compared with the prior art, the above-described technical solutions conceived in this invention can achieve the following results.

[0037] Beneficial effects:

[0038] (1) Adaptable to low computing power scenarios. Extreme learning machine models have low computational complexity and low requirements for device computing power, making them suitable for intelligent edge scenarios with strong computing resource constraints.

[0039] (2) Adaptable to low sampling frequency scenarios. Extreme learning machines can achieve high-precision learning under small sample data conditions, which can effectively cope with the data scarcity problem caused by low-frequency sampling of industrial equipment.

[0040] (3) Accurate fault diagnosis. The powerful nonlinear simulation capability of the Extreme Learning Machine model can accurately learn the characteristics of various soft faults in the operation of the DC-DC converter and effectively diagnose fault types.

[0041] (4) Optimize diagnostic efficiency. Based on the DC-DC converter output voltage ripple formula, the cause of the fault can be quickly located through parameter matching, the reliability optimization strategy can be clarified, and the fault diagnosis can be automated.

[0042] (5) Precise reliability optimization. Based on the diagnostic results, matching optimization schemes are adopted for different fault types, such as adjusting the switching frequency, to effectively improve the operational reliability of the DC-DC converter.

[0043] (6) Real-time decision support. Joint diagnostics and optimization at the edge can respond in real time to changes in the operating status of the DC-DC converter, providing support for decision-making. Attached Figure Description

[0044] Figure 1 The flowchart of the degradation fault output voltage ripple optimization method for DC-DC converters provided by the present invention is shown.

[0045] Figure 2 This is a schematic diagram of an SLFN network.

[0046] Figure 3 This is the circuit diagram for the BUCK converter.

[0047] Figure 4 The results show the diagnosis and switching frequency adjustment for different degradation conditions in the examples. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0049] To address the limitations of low computing power and low sampling frequency, the diagnostic method proposed in this invention mainly includes the following features: It requires analysis of the output voltage V. i (t) Undersampled signal acquisition is performed to adapt to low sampling frequency conditions; signal time-domain features are extracted; a feature selection method based on Support Vector Machine Recursive Feature Elimination (SVM-RFE) is used to select the optimal combination of as few features as possible, thereby reducing the size and computational requirements of the diagnostic model; extreme learning machine (ELM) based on Bayesian optimization is used for fault diagnosis, and the model is optimized and trained on a PC with sufficient computing resources using a computationally complex optimization algorithm, and finally the trained model parameters are deployed on a DSP; the relationship between the degradation of key components and output voltage ripple is analyzed to determine appropriate optimization strategies, and an attempt is made to optimize the control of the DC-DC converter under the degradation conditions of diagnosis. The main process of the proposed diagnostic and optimization method is as follows: Figure 1 As shown.

[0050] Due to the characteristics of periodic signals, the signal strength at the same time point is the same across different periods. Utilizing this characteristic, sampling points originally concentrated in one period can be distributed across multiple periods. The sampling time points are different for each period. The sampling results are stored in corresponding positions in an array according to certain rules. Once all sampling results are stored, the effect of obtaining a high-frequency sampling signal can be achieved using a low-frequency sampling signal. Therefore, considering the high switching frequency of the DC-DC converter and real-time performance, a sampling strategy of acquiring 4 points over 5 periods is adopted. The sampling interval at this time is:

[0051]

[0052] Based on the switching frequency f of the selected DC-DC converter, the sampling frequency can be obtained as follows:

[0053]

[0054] Seventeen typical time-domain features of the output voltage ripple were extracted to form a feature dataset: Feature extraction was performed on the processed sampled data. Considering the uncertainty of the switching frequency, the extraction of signal frequency domain features and wavelet features was abandoned, and only time-domain features were used for fault diagnosis. Seventeen time-domain features were extracted, including maximum value, minimum value, peak value, peak-to-peak value, average value, absolute average value, square root amplitude, variance, standard deviation, root mean square, kurtosis, skewness, shape factor, peak factor, impulse factor, margin factor, and gap coefficient.

[0055] However, too many features increase computational cost and may reduce the diagnostic accuracy of the model. Therefore, SVM-RFE is adopted, using a backward strategy to recursively eliminate the 17 extracted temporal features. A Gaussian kernel-based SVM is used as the base model for multi-round training. After each round of training, the features with the lowest weights (feature weight coefficients or feature importance) are eliminated. Then, the next round of training and elimination is performed based on the updated feature set, continuing until the required number of most important features remain.

[0056] SVM-RFE is a sequential backward selection algorithm based on the maximum margin principle of SVM. It trains the model with samples and then scores and ranks each temporal feature. It removes the temporal feature with the lowest score and then retrains the model using the remaining temporal features for the next iteration. Finally, it selects the desired number of optimal features. The ranking criterion score of temporal feature i is defined as:

[0057]

[0058] Where i = 1, 2, ..., n, ω i denoted as the weight of time-domain feature i.

[0059] The basic model chosen for fault diagnosis is ELM. A key feature of ELM is that the parameters of the hidden layer nodes are randomly generated, resulting in very fast model training. From a network structure perspective, ELM is a single-layer feedforward neural network (SLFN), such as... Figure 2 As shown.

[0060] This SLFN consists of three layers: an input layer, hidden layers, and an output layer (or a two-layer network if the input layer is ignored). The hidden layer contains L... E The neural network has N hidden neurons, N input layer nodes, and an m-dimensional output vector (which is one-dimensional for binary classification). Ignoring the input and hidden layers and considering only the outputs of the hidden neurons and the output layer, the output function of the neural network for the training sample x can be expressed as:

[0061]

[0062] Where a i and b i It is a parameter of the hidden layer node, β i Let G represent the connection weights between the i-th hidden neuron and the output neuron, which is an m-dimensional weight vector. In the formula, G represents the output of the hidden layer neuron. For additional hidden neurons, G is:

[0063] G(a i ,b i ,x)=g(a i ·x+b i )

[0064] Where g is the activation function, which can be a linear function, a sigmoid function, etc.; for RBF hidden neurons, G is:

[0065] G(a i ,b i ,x)=g(b i ||a i -x||)

[0066] Where a i and b i Let represent the center and influence factor of the i-th radial basis function node, respectively.

[0067] The output of a neural network can also be represented as:

[0068] Hβ=T

[0069] Where H is the result matrix of the hidden layer, β is the weight matrix of the output layer, and T is the desired output matrix.

[0070] If the neural network can accurately predict the training samples, then the weights of the hidden and output layers are solvable. Especially when L...E When L = N, there is definitely a solution. However, in practical problems, L... E This value is often much smaller than N. Therefore, the problem of finding the weight vector becomes unsolvable, meaning there is an error between the network output and the actual value. The cost function that measures the error between the network output and the actual value can be defined as:

[0071] J = (Hβ - T) T (Hβ-T)

[0072] Next, we need to find the optimal weight vector that minimizes the cost function J. ELM solves this problem in two cases:

[0073] 1) If H has full column rank, the optimal weights can be found using the least squares method, and the solution is:

[0074]

[0075] in:

[0076] H * =(H TH ) -1 H T

[0077] 2) If H does not have full column rank, the optimal weights are calculated by obtaining the pseudo-inverse of H through singular value decomposition.

[0078] Unlike Backpropagation Neural Networks (BPNNs), which use gradient descent to iteratively update the weights across all layers, ELM does not adjust the input and hidden layer weights of SLFNs. These weights are randomly assigned, making ELM training very fast. ELM focuses on using ordinary least squares to determine the weights from the hidden layers to the output layer.

[0079] Based on the ELM classification calculation process, the ELM model is directly reconstructed in the DSP for fault diagnosis. The calculation process for output Y, given input sample data X, is as follows:

[0080] Y = H·β

[0081] in:

[0082] H = sofplus(X·ω+b)

[0083] Here, softplus is the activation function, and H is the result of the hidden layer.

[0084] During ELM training, the parameters ω and bias b of the hidden layer nodes are first randomly initialized. For the training dataset X, the result of the hidden layer H = sofplus(X·ω+b) is calculated. Then, the result of the output layer is calculated according to Y = H·β. Finally, the cost function J = (Hβ-T) is used.T (Hβ-T) updates the model's parameter β using the least squares method, and the model training is complete when the cost function is reduced to an acceptable level.

[0085] However, the choice of network hyperparameters significantly impacts the final performance. Compared to traditional black-box optimization methods such as grid search and random search, Bayesian optimization has become the mainstream approach for addressing these challenges due to its high data efficiency. This is attributed to its ability to incorporate prior beliefs about the problem to guide the sampling of new data and to achieve a good balance between exploration and exploitation during the search process.

[0086] Bayesian optimization is a method that uses Bayesian theory to guide parameter search to find the minimum or maximum value of an objective function. In each iteration, it utilizes previously observed historical information (prior knowledge) to guide the next optimization step, thus greatly improving the efficiency of parameter search.

[0087] Bayesian optimization utilizes Bayesian theory to construct the posterior distribution of the objective function. It uses this posterior distribution to determine the promising hyperparameters to be evaluated next. Typically, a Gaussian process is used to model the posterior distribution and is updated after each evaluation using Bayesian inference. The obtained function (usually the expected improvement) determines the hyperparameters to be evaluated next based on the posterior distribution. By carefully balancing exploration and exploitation in its recommendations, Bayesian optimization can find optimal hyperparameters more efficiently than methods such as grid search or random search. Overall, by learning from previous evaluations, Bayesian optimization can more intelligently guide the search.

[0088] Specifically, Bayesian optimization is performed as follows:

[0089] Step 1: Initialize the Gaussian process with prior knowledge based on any available prior knowledge.

[0090] Step 2: Use the acquisition function to determine the next sampling point based on the current Gaussian process posterior distribution. A common acquisition function is the expected improvement.

[0091] Step 3: Evaluate the objective function at the selected point and update the Gaussian process to obtain a new posterior distribution.

[0092] Step 4: Repeat steps 2-3, using the updated posterior as the new prior for the Gaussian process.

[0093] Therefore, by iteratively sampling and updating the probabilistic model at intelligently selected points, Bayesian optimization can effectively find the minimum or maximum value of an unknown objective function. The surrogate function assists in optimization by understanding the objective probabilistically.

[0094] The selected ELM model was optimized using Bayesian optimization, which utilized a Gaussian process and an EI acquisition function. The Bayesian optimization algorithm evaluated the training results of the ELM model on the training set data under different hyperparameter combinations and searched for potential optimal hyperparameter combinations based on the known results. The F1 score of the ELM model's diagnostic results (F1 = 2 * (precision * recall) / (precision + recall), where precision = TP / (TP + FP), recall = TP / (TP + FN), where TP (TruePositive) represents the number of true positives, FP (False Positive) represents the number of negatives incorrectly predicted as positives, and FN (False Negative) represents the number of positives incorrectly predicted as negatives) was used as the objective function value for Bayesian optimization, which was run for 100 iterations. After Bayesian hyperparameter optimization, the selected parameters for ELM were C. E =5.27 and L E =213. Parameter C E Controls the strength of the applied regularization. Parameter L E This indicates the number of nodes in the hidden layer.

[0095] In a DC-DC converter, the formula for calculating the peak-to-peak value of the output voltage ripple is as follows:

[0096]

[0097] Where f is the switching frequency, L is the filter inductance, C is the output capacitor, and V is the voltage. in V is the input voltage. o This is the output voltage.

[0098] Therefore, the magnitude of the output voltage ripple is inversely proportional to the switching frequency f and the product of L and C. Thus, when L and C deteriorate and their values ​​decrease, the switching frequency f needs to be increased to reduce the output voltage ripple.

[0099] According to the formula, the relationship between the switching frequency f and L and C can be obtained as follows:

[0100]

[0101] If the degradation parameters L' and C' are known, then their respective degradation ratios are a. l =L' / L, a c =C' / C. Assuming the ripple amplitude remains constant by increasing the switching frequency, the adjusted frequency should be f' = f × a. l ×a c .

[0102] The corresponding switching frequency adjustment level and diagnostic results table are written into the DSP. Based on the diagnostic results, the switching frequency adjustment table can be queried, and the switching frequency that needs to be adjusted can be output to improve the reliability of the converter.

[0103] The Buck converter is used as the implementation target. Its structure is as follows: Figure 3 As shown, where V in and V o These are the input and output voltages, C. in The input filter capacitor is S1. S2 are two power switches, and R is the load resistor. The output filter inductor is modeled with resistor R. L An inductor L is connected in series. The output capacitor is modeled as being connected to a resistor R. C A capacitor C is connected in series. The converter parameters are shown in Table 1. The normal and abnormal groups of the Buck converter are shown in Table 2. The tolerances for the inductor and capacitor are set to 5%. When the attenuation of L and C is within 0-4%, the converter is considered normal; when the degradation of L is within 5-9%, the inductor is considered slightly degraded; when the degradation of L is within 10-14%, the inductor is considered moderately degraded; and when the degradation of L is within 15-19%, the inductor is considered severely degraded. The degradation interval of capacitor C is the same. Table 2 shows 16 fault types. By changing the capacitance and inductance parameters, normal and various soft fault states are simulated. Normal state and 15 abnormal states are given. For each state, there are 5 parameter settings for L and C, for a total of 25 parameter combinations and 16 states. One set of data is collected for each parameter combination, for a total of 400 sets of experimental data. Then, the data is undersampled to simulate the low sampling frequency under actual conditions. The initial switching frequency was 50 kHz, the sampling frequency was 40 kHz, and data was collected for 10 cycles, with 8 sampling points per cycle. Due to the limited computing power of the DSP, the diagnostic model was first trained on a PC using the collected dataset. Then, the parameters ω, b, and beta (representing the random input layer weights and biases) of the trained ELM model and the trained output layer weight matrix were retrieved.

[0104] Table 1

[0105]

[0106] After processing, it is 4×10 4Feature extraction was performed on the Hz sampled data. Considering the uncertainty of the switching frequency, the extraction of signal frequency domain features and wavelet features was abandoned, and only time domain features were used for fault diagnosis. Seventeen time domain features were extracted, including maximum value, minimum value, peak value, peak-to-peak value, average value, absolute average value, square root amplitude, variance, standard deviation, root mean square, kurtosis, skewness, shape factor, peak factor, impulse factor, margin factor, and gap coefficient. However, too many features would increase the computational load and reduce the diagnostic speed of the model. Therefore, SVM-RFE was used to perform recursive feature elimination on the 17 extracted time domain features to obtain an appropriate number of features. SVM was used as the base model for multi-round training. After each round of training, the feature with the lowest weight (feature weight coefficient or feature importance) was eliminated, and the next round of training and elimination was carried out based on the new feature set until the three most important features remained.

[0107] Table 2

[0108]

[0109] The feature selection data was divided into test and training sets in a 3:7 ratio. A Bayesian optimization algorithm employing a Gaussian process and Expectation Improvement (EI) acquisition function was used, combined with the training set data, to optimize the ELM model hyperparameter C through multiple iterative training cycles. E L E This allows us to find the optimal combination of hyperparameters. Parameter C E Primarily used as a regularization term for ill-posed linear problems, regularization is a technique used to prevent overfitting in machine learning models and improve generalization performance. Parameter C E Controls the strength of the applied regularization. Parameter L E This represents the number of nodes in the hidden layer. Bayesian optimization ran for 100 iterations (n_iter = 100). The ELM activation function was chosen as the softplus method, combined with a regularization strategy. The One-vs-rest strategy assigns one class of data as class 1 (positive) and the remaining m-1 classes of data as class 2 (negative), thus training m binary classifiers. After Bayesian hyperparameter optimization, the selected parameter C... E =5.27, L E =213, and the elements of the weight matrix and bias matrix are uniformly and randomly generated in the interval (-1,1). For the experimental dataset, the model performance is shown in Table 3, with an average diagnostic accuracy of 95.4% for 16 degradation conditions.

[0110] Table 3 Simulation Results

[0111]

[0112] As can be seen, the model achieves robust diagnostic performance for all 16 soft faults. Based on the diagnostic results, estimated degradation rates for L and C can be provided. By using the median percentage degradation of L and C in each fault category as its estimated percentage degradation, the corresponding switching frequency adjustment table can be calculated.

[0113] According to the ELM classification calculation process, the parameter matrix and related code are written into the DSP, and the ELM model is directly reconstructed in the DSP for fault diagnosis. The adjustment level corresponding to the switching frequency and the diagnostic results are also written into the DSP. Based on the diagnostic results, the switching frequency adjustment table can be queried, and the required switching frequency can be output. Single output voltage data of the sensor under various operating conditions are collected. After data processing, samples at low sampling frequencies are sent to the DSP for soft fault diagnosis based on a ten-cycle, eight-point ELM model. After 100 diagnoses, the overall diagnostic accuracy on the DSP is 0.95, similar to the results on the PC. The code file size is 103kb, with 375 lines. The stored matrix is: beta=[213,16], ω=[3,213], b=[1,213]. The average single diagnosis time is close to 1.6ms. Furthermore, the DSP can output the correct switching frequency while correctly identifying the fault. Figure 4 The output voltage ripple waveform of the BUCK converter after diagnostics by the DSP and switching frequency adjustment is shown.

[0114] The results of the implementation examples verify the feasibility of deploying and applying the proposed method under conditions of low computing power and low sampling frequency. It can perform relatively effective soft fault diagnosis of DC-DC converters in DSPs and provide control optimization strategies based on the results.

[0115] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing the degradation fault output voltage ripple of a DC-DC converter, characterized in that, Includes the following steps: S1. Undersample the output voltage of the DC-DC converter to obtain the output voltage ripple; S2. Extract the time-domain features of the output voltage ripple to form a feature dataset; S3. Feature selection is performed using recursive feature elimination based on support vector machines (SVMs) to select the optimal number of optimal features, forming an optimal feature dataset. The recursive feature elimination based on SVMs includes: establishing a sample training model, scoring and ranking each temporal feature, deleting the temporal feature with the lowest feature score, and then retraining the model using the remaining temporal features for the next iteration until the required number of optimal features are selected; wherein the temporal features... i The ranking criteria score is defined as follows: in, i =1,2,…n, Time-domain features i The weights; S4. The fault diagnosis model is trained using the optimal feature dataset, which is based on an Extreme Learning Machine (ELM). The hyperparameters of the fault diagnosis model are optimized through multiple iterative training cycles to obtain the optimal combination of hyperparameters, thus obtaining a trained fault diagnosis model. The training steps of the fault diagnosis model are as follows: An ELM network is selected as the basic network model for fault diagnosis. First, the network structure of the ELM is randomly initialized. A portion of the optimal feature dataset is used as the training set to train the ELM network parameters. During training, the model gradually adjusts the network parameters using the least squares method to better fit the training data. Then, the remaining optimal feature dataset is used as the validation set to verify the diagnostic accuracy under this network structure. Based on the validation results of the validation set, the ELM network hyperparameters that best suit the fault diagnosis task dataset are found through continuous iterative optimization using a Bayesian optimization algorithm; that is, the optimal network structure is found. Finally, based on the optimal network structure, the training set is used to train the model, resulting in a trained fault diagnosis model. The Bayesian optimization algorithm includes: Step 1: Initialize the Gaussian process with prior knowledge based on any available prior knowledge; Step 2: Use the acquisition function to determine the next sampling point based on the current Gaussian process posterior distribution; Step 3: Evaluate the objective function at the selected point and update the Gaussian process to obtain a new posterior distribution; Step 4: Repeat steps 2-3, using the updated posterior as the new prior for the Gaussian process; S5. Input the real-time output voltage of the DC-DC converter to the trained fault diagnosis model to perform real-time diagnosis of the DC-DC converter; S6. Based on the real-time diagnostic results of the DC-DC converter, infer the degree of degradation of the capacitors and inductors of the DC-DC converter, adjust the switching frequency, and reduce the output voltage ripple; the peak-to-peak value of the output voltage ripple of the DC-DC converter is expressed as: in, f Where L is the switching frequency, L is the filter inductance, and C is the output capacitor. V in Input voltage, V o This is the output voltage.

2. The method according to claim 1, characterized in that, The time-domain features include maximum value, minimum value, peak value, peak-to-peak value, average value, absolute average value, square root amplitude, variance, standard deviation, root mean square, kurtosis, skewness, shape factor, peak factor, impulse factor, margin factor, and gap coefficient.

3. The method according to claim 1, characterized in that, The undersampling frequency is set to f ,in, For any positive integer, f This refers to the switching frequency of the DC-DC converter.

4. A degradation fault output voltage ripple optimization system for a DC-DC converter, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the degradation fault output voltage ripple optimization method for a DC-DC converter as described in any one of claims 1 to 3.

Images