A method for predicting the life of a high-frequency transformer
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-24
- Publication Date
- 2026-08-11
AI Technical Summary
[0005]针对现有技术中存在的缺陷，本发明的目的在于提供一种高频变压器的寿命预测方法，解决了小样本下的高频变压器的健康状态难预测的问题
1、构建了异构特征编码器，在全局退化趋势感知与局部波动捕捉之间实现了卓越的平衡，克服了传统单一尺度模型的局限性；
Smart Images
Figure CN122548679A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of isolated high-frequency power conversion technology in power electronics, specifically relating to a method for predicting the lifespan of a high-frequency transformer. Background Technology
[0002] Isolated DC-DC converters, with their superior voltage regulation performance and reliable electrical isolation capabilities, have played an irreplaceable role in critical fields such as data centers, electric vehicles, and solid-state transformers. These applications place higher demands on the power density and operating efficiency of energy conversion units. With the continuous improvement of system integration, system reliability assessment is expanding from single components to key components throughout the entire value chain. As a core component of isolated DC-DC converters, the high-frequency transformer (HFT) is receiving unprecedented attention from academia and industry regarding its operational lifespan and reliability under high-voltage, high-power conditions.
[0003] High-frequency transformers (HFTs) differ significantly from traditional power frequency transformers in their operating mechanisms and failure modes. Power frequency transformers operate in a sinusoidal environment at low frequencies for extended periods, resulting in a relatively gradual degradation process. In contrast, HFTs are deployed in high-frequency conversion circuits and are subjected to high degradation rates generated by power semiconductor devices. v / d t Pulse-width modulated voltage excitation not only leads to severe skin and proximity effects, but also causes intense nonlinear losses and electro-thermal-magnetic multi-field coupling stress in the core and windings during operation. Professor Shen Zhan's team at Southeast University conducted the first accelerated thermal aging test observation of high-frequency transformers. The study placed experimental samples under high-intensity thermal stress conditions of 180℃ and 200℃ for long-term observation and provided detailed degradation test results for key parameters. Based on the experimental findings, the failure of magnetic components is not an instantaneous physical destruction, but a gradual performance degradation process, ultimately manifesting as the loss of energy transmission function due to insulation breakdown, core failure, winding short circuit, or thermal runaway. The remaining lifetime at 140℃ was derived using the Arrhenius equation and Weibull probability distribution, which describe the effect of temperature on material aging. However, a single empirical formula is insufficient to accurately predict the nonlinear degradation process.
[0004] In recent years, data-driven methods have attracted widespread attention in the field of lifetime prediction. The lifetime prediction of High-Fold Fiber (HFT) is logically consistent with typical time series problems such as lithium battery lifetime prediction, power load prediction, and wind turbine remaining lifetime prediction, all exhibiting characteristics of long-term degradation trends and short-term nonlinear fluctuations in key parameters over time. Compared to traditional empirical formulas that rely on prior knowledge, data-driven methods do not require in-depth exploration of complex internal physicochemical degradation processes and can adaptively capture the changing patterns of target parameters by mining measurement data. However, existing time series models still have limitations in handling HFT lifetime prediction. While Recurrent Neural Networks (RNNs) and their variants LSTM and GRU can handle sequential data, their recursive computation mechanism leads to slow inference speeds, and when faced with ultra-long time series data, they are limited by the gradient vanishing problem, making it difficult to effectively capture historical degradation information over long distances. Convolutional Neural Networks (CNNs), while improving efficiency through parallel computation, are limited by the receptive field of the convolutional kernels, making it difficult to capture long-term feature correlations. For HFT, the State of Health (SOH) encompasses both global aging trends and local fluctuations, making it difficult for a single model to simultaneously account for both. The Transformer architecture, with its self-attention mechanism, possesses powerful global modeling capabilities, but its computational complexity increases quadratically with sequence length. O ( L 2 Furthermore, CNNs are highly prone to overfitting on small sample data generated during HFT aging experiments. In summary, CNNs struggle to capture long-term feature relationships, but can perform parallel computations. RNNs theoretically possess the ability to model long-term dependencies, but in practice, their memory windows are very short. While LSTM and GRU address some of the vanishing gradient problem, their memory windows remain limited. Attention mechanisms can capture global information, but their computational complexity is extremely high when processing time-series data, limiting the scalability of models handling long sequence inputs. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method for predicting the lifespan of high-frequency transformers, solving the problem of difficulty in predicting the health status of high-frequency transformers under small sample conditions. Using the method described in this invention, the prediction accuracy and stability of the model in small sample environments can be effectively improved.
[0006] To achieve the above objectives, the present invention provides a method for predicting the lifespan of a high-frequency transformer, comprising the following steps: S1. Preprocess the measurement data of the high-frequency transformer; S2. Define the lifespan standard for high-frequency transformers; S3. Utilize heterogeneous feature encoders to extract global aging trends and multi-scale local fluctuation features; S4. Utilize the XGBoost multi-output forest regressor to achieve better generalization ability and anti-overfitting characteristics.
[0007] Further preferably, the high-frequency transformer measurement data is preprocessed, including: The high-frequency transformers were divided into 24 groups. Groups 1-12 underwent thermal aging tests at 200℃, while groups 13-24 underwent thermal aging tests at 180℃. During the experiments, primary inductance, primary resistance, leakage inductance, and short-circuit resistance were measured. Primary inductance represents the effective equivalent permeability of the core, while primary resistance represents core loss and winding loss. Under 200℃ conditions, due to differences in manufacturing processes and core materials, the measured parameters of the 12 groups showed some dispersion, but the core aging trend was highly consistent. To quantify the data variation trend, the data from the 12 groups were averaged and normalized using the following formula: Specifically, the primary inductance of all samples showed a non-linear decreasing trend with aging time, with a faster decay rate in the early stage and a gradual flattening in the later stage; the primary resistance showed the same changing trend as the primary inductance; the leakage inductance showed a basically unchanged trend with aging time; while the short-circuit resistance showed a monotonically increasing characteristic.
[0008] Further preferred, defining lifespan standards for high-frequency transformers; including: The failure rate of high-frequency transformers in actual converter operation is much lower than that of capacitors. However, with the continuous increase in converter voltage levels and power density, and the stringent limitations on the design margin of high-frequency transformers for commercial applications, the life prediction of high-frequency transformers is facing increasingly severe challenges. Therefore, this paper studies the life prediction of high-frequency transformers. Following the life-end criterion of capacitors, namely, the capacitance value decreasing to 80% of the initial capacity or the ESR increasing to 2-3 times the initial value, we define the life-end criterion of high-frequency transformers as the state of equilibrium (SOH) decreasing to 80% of the initial capacity. The SOH of high-frequency transformers is defined as follows: .
[0009] Further optimized, a heterogeneous feature encoder is used to extract global aging trends and multi-scale local fluctuation features: Unlike Informer's encoder, the heterogeneous feature encoder employs a heterogeneous lateral stacking strategy to capture changing trends across different time scales. First, it utilizes the ProbSparse self-attention mechanism to model long-range temporal dependencies and global degradation trends, with its input being... ,L Given the length of the input sequence, d The dimension of the input; q i for The first in the matrix i A sequence, k j for The first in the matrix j A sequence, due to The dot product values exhibit a long-tailed distribution, therefore the Top-order selection method is utilized. u An active q j Formed with q j Sparse matrices of the same size : Each key only focuses on Top- u An active q j After passing through the softmax function, multiplying by the value yields ProbSparse self-attention. If you choose u For log L For a given number of samples, the computational complexity of the ProbSparse attention mechanism can be reduced from... O ( L 2 Reduced to O ( L log L To collaboratively model the global degradation trend and local fluctuation characteristics of the SOH of the high-frequency transformer, a temporal convolutional network was added as a heterogeneous branch to the encoder, effectively compensating for the limitations of the ProbSparse self-attention mechanism in capturing local fluctuations; the dilated convolution in the TCN is represented by the following equation: The residual block structure introduced on the output side enables each layer of the network to learn corrections to the identity mapping. This has been repeatedly proven to be helpful for training very deep networks, as shown in the equation: .
[0010] Further optimization utilizes the XGBoost multi-output forest regressor to achieve better generalization ability and anti-overfitting characteristics.
[0011] Choosing XGBoost instead of traditional generative decoders has the following advantages: First, XGBoost is a non-data-intensive algorithm, meaning that satisfactory regression results can be obtained with only a small amount of data; second, XGBoost consists of a series of weak decision trees, summarizing the training results of all weak evaluators to achieve better regression performance, thus mitigating overfitting in scenarios with limited experimental samples; third, XGBoost has a simple structure and its hyperparameters are easy to adjust; i Use decision trees to learn target values y i and i -1 residuals from the sum of the outputs of the decision trees; using a boosting learning strategy, in generating the... k When planting trees, we need to minimize the objective function of the following expression. obj To adjust the first k Tree learning parameters θ When the objective function is minimized, the output value is... f k close to y k The final output can be obtained by summing the output values of all decision trees. To achieve the future N The health status prediction step employs a multi-output forest regressor strategy. This is achieved by establishing... N Each XGBoost model outputs a single SOH step, thereby achieving the goal of multiple outputs.
[0012] The high-frequency transformer lifetime prediction method disclosed in this application has the following advantages compared with the prior art: 1. A heterogeneous feature encoder was constructed, which achieved an excellent balance between global degradation trend perception and local fluctuation capture, overcoming the limitations of traditional single-scale models; 2. The innovative introduction of the XGBoost multi-output forest regressor demonstrates that this approach is superior to the traditional generative decoder, exhibiting stronger robustness and prediction accuracy on the limited sample dataset generated by the HFT aging experiment. Attached Figure Description
[0013] Figure 1 A flowchart illustrating the steps of a method for predicting the lifespan of a high-frequency transformer; Figure 2 Parameter measurement methods and equivalent circuit diagrams; Figure 3 Graph showing the relationship between measured parameters and aging time; Figure 4Relationship between SOH and aging time of high-frequency transformer; Figure 5 NeuroForest-Former model diagram; Figure 6 ProbSparse self-attention mechanism diagram; Figure 7 TCN structure diagram; Figure 8 XGboost model diagram; Figure 9 Comparison of prediction results from different time series models; Detailed Implementation
[0014] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0015] like Figure 1 As shown, the present invention provides a method for predicting the lifespan of a high-frequency transformer, comprising the following steps: S1. Preprocess the measurement data of the high-frequency transformer; The high-frequency transformers were divided into 24 groups. Groups 1-12 underwent thermal aging tests at 200℃, while groups 13-24 underwent thermal aging tests at 180℃. During the experiments, the primary inductance, primary resistance, leakage inductance, and short-circuit resistance were measured. The primary inductance represents the effective equivalent permeability of the core, while the primary resistance represents core loss and winding loss. Figure 2 shows the specific measurement methods and corresponding equivalent circuit diagrams for these parameters. Under 200℃ conditions, due to differences in manufacturing processes and core materials, the measured parameters of the 12 groups exhibited some dispersion, but the core aging trend showed a high degree of consistency. To quantify the data variation trend, the 12 groups of data were averaged and normalized using the following formula: The relationship between the measured data and aging time is shown in Figure 3. Specifically, the primary inductance of all samples shows a non-linear decreasing trend with aging time, with a faster decay rate in the early stage and a gradual flattening in the later stage; the primary resistance has the same changing trend as the primary inductance; the leakage inductance shows a basically unchanged trend with aging time; while the short-circuit resistance shows a monotonically increasing characteristic.
[0016] S2. Define the lifespan standard for high-frequency transformers; Specifically as follows: The failure rate of high-frequency transformers in actual converter operation is much lower than that of capacitors. However, with the continuous increase in converter voltage levels and power density, and the stringent limitations on the design margin of high-frequency transformers for commercial applications, the life prediction of high-frequency transformers is facing increasingly severe challenges. Therefore, this paper studies the life prediction of high-frequency transformers. Following the life-end criterion of capacitors, namely, the capacitance value decreasing to 80% of the initial capacity or the ESR increasing to 2 to 3 times the initial value, we define the life-end criterion of high-frequency transformers as the SOH decreasing to 80% of the initial capacity. The SOH of high-frequency transformers is defined as shown in the equation. The SOH of high-frequency transformers is calculated using the equation, as shown in Figure 4. S3. Utilize heterogeneous feature encoders to extract global aging trends and multi-scale local fluctuation features; The proposed NeuroForest-Former employs a decoupled two-stage architecture: a heterogeneous feature encoder and an XGBoost multi-output forest regressor, as shown in Figure 5. Unlike the Informer encoder, the heterogeneous feature encoder uses a heterogeneous lateral stacking strategy to capture trends across different time scales. First, it utilizes the ProbSparse self-attention mechanism to model long-range temporal dependencies and global degradation trends, as shown in Figure 6. Its input is... , L Given the length of the input sequence, d The dimension is the input dimension. q i for The first in the matrix i A sequence, k j for The first in the matrix j A sequence, due to The dot product values exhibit a long-tailed distribution, therefore the Top-order selection method is utilized. u An active q j Formed with q j Sparse matrices of the same size . Each key only focuses on Top- u An active q j After passing through the softmax function, multiplying by the value yields ProbSparse self-attention. If you choose u For log LFor a given number of samples, the computational complexity of the ProbSparse attention mechanism can be reduced from... O ( L 2 Reduced to O ( L log L To collaboratively model the global degradation trend and local fluctuation characteristics of the SOH (Self-Attention) transformer, a temporal convolutional network (TCN) was added as a heterogeneous branch to the encoder, as shown in Figure 7(a). This effectively compensates for the limitations of the ProbSparse self-attention mechanism in capturing local fluctuations. The dilated convolution in the TCN is represented by the following equation: Introduced on the output side, such as Figure 7 The residual block structure shown in (b) enables each layer of the network to learn corrections to the identity mapping, which has been repeatedly proven to be helpful for training very deep networks, as shown in the equation: In the experiment, sampled values of the output current under light, medium, and heavy loads were extracted, as shown in Figure 3. To quantify the proportion of output current noise under different power levels, the relative peak-to-peak value of the output current was calculated using the following formula: S4. Utilize the XGBoost multi-output forest regressor to achieve better generalization ability and anti-overfitting characteristics.
[0017] Choosing XGBoost instead of traditional generative decoders has the following advantages. First, XGBoost is a non-data-intensive algorithm, meaning that satisfactory regression results can be obtained with only a small amount of data. Second, XGBoost consists of a series of weak decision trees, summarizing the training results of all weak evaluators to achieve better regression performance, thus mitigating overfitting in scenarios with limited experimental samples. Third, XGBoost has a simple structure, and its hyperparameters are easy to adjust. The training process of XGBoost is shown in Figure 8. i Use decision trees to learn target values y i and i The residual of the sum of the outputs of -1 decision trees. Using a boosting learning strategy, in generating the... k When planting trees, we need to minimize the objective function of the following expression. obj To adjust the first k Tree learning parameters θ When the objective function is minimized, the output value is... f k close to y kThe final output can be obtained by summing the output values of all decision trees. To achieve the future N The health status prediction step employs a multi-output forest regressor strategy. This is achieved by establishing... N Each XGBoost model outputs a single SOH step, thereby achieving the goal of multiple outputs.
[0018] To verify the effectiveness of the proposed NeuroForest-Former model, it was compared with advanced time series models such as Informer, Informer(stack), Transformer, Transformer(Encoder), RNN, LSTM, and GRU. Figure 9 shows the prediction plots of different time series models.
[0019] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. A method for predicting the lifespan of a high-frequency transformer, characterized in that, Includes the following steps: S1. Preprocess the measurement data of the high-frequency transformer; S2. Define the lifespan standard for high-frequency transformers; S3. Utilize heterogeneous feature encoders to extract global aging trends and multi-scale local fluctuation features; S4. Utilize the XGBoost multi-output forest regressor to achieve better generalization ability and anti-overfitting characteristics.
2. The method for predicting the lifespan of a high-frequency transformer as described in claim 1, characterized in that, In S1, the high-frequency transformers were divided into 24 groups. Groups 1-12 underwent thermal aging tests at 200℃, while groups 13-24 underwent thermal aging tests at 180℃. During the experiments, the primary inductance, primary resistance, leakage inductance, and short-circuit resistance were measured. The primary inductance represents the effective equivalent permeability of the core, while the primary resistance represents core loss and winding loss. Under 200℃ conditions, due to differences in manufacturing processes and core materials, the measured parameters of the 12 groups exhibited some dispersion, but the core aging trend showed a high degree of consistency. To quantify the data variation trend, the 12 groups of data were averaged and normalized using the following formula: Specifically, the primary inductance of all samples showed a non-linear decreasing trend with aging time, with a faster decay rate in the early stage and a gradual flattening in the later stage; the primary resistance showed the same changing trend as the primary inductance; the leakage inductance showed a basically unchanged trend with aging time; while the short-circuit resistance showed a monotonically increasing characteristic.
3. The method for predicting the lifespan of a high-frequency transformer as described in claim 1, characterized in that, In S2, the failure rate of high-frequency transformers during actual converter operation is much lower than that of capacitors. However, with the continuous increase in converter voltage levels and power density, and the stringent limitations on the design margin of high-frequency transformers for commercial applications, the life prediction of high-frequency transformers is facing increasingly severe challenges. Therefore, the life prediction of high-frequency transformers has been studied. Following the life end criterion of capacitors, namely, the capacitance value decreases to 80% of the initial capacity or the ESR increases to 2 to 3 times the initial value, we define the life end criterion of high-frequency transformers as the SOH decreasing to 80% of the initial capacity. The SOH of high-frequency transformers is defined as follows: 。 4. The method for predicting the lifespan of a high-frequency transformer as described in claim 1, characterized in that, In S3, the heterogeneous feature encoder differs from the Informer encoder, employing a heterogeneous lateral stacking strategy to capture changing trends across different time scales. It first utilizes the ProbSparse self-attention mechanism to model long-range temporal dependencies and global degradation trends, with its input being... , L Given the length of the input sequence, d The dimension of the input; q i for The first in the matrix i A sequence, k j for The first in the matrix j A sequence, due to The dot product values exhibit a long-tailed distribution, therefore the Top-order selection method is utilized. u An active q j Formed with q j Sparse matrices of the same size : Each key only focuses on Top- u An active q j After passing through the softmax function, multiplying by the value yields ProbSparse self-attention. If you choose u For log L For a given number of samples, the computational complexity of the ProbSparse attention mechanism can be reduced from... O ( L 2 Reduced to O ( L log L To collaboratively model the global degradation trend and local fluctuation characteristics of the SOH of the high-frequency transformer, a temporal convolutional network was added as a heterogeneous branch to the encoder, effectively compensating for the limitations of the ProbSparse self-attention mechanism in capturing local fluctuations; the dilated convolution in the TCN is represented by the following equation: The residual block structure introduced on the output side enables each layer of the network to learn corrections to the identity mapping. This has been repeatedly proven to be helpful for training very deep networks, as shown in the equation: 。 5. The method for predicting the lifespan of a high-frequency transformer as described in claim 1, characterized in that, In S4, choosing XGBoost instead of the traditional generative decoder has the following advantages: First, XGBoost is a non-data-intensive algorithm, meaning that only a small amount of data is needed to achieve satisfactory regression results; second, XGBoost consists of a series of weak decision trees, summarizing the training results of all weak evaluators to achieve better regression performance, thus having the advantage of mitigating overfitting in scenarios with limited experimental samples; third, XGBoost has a simple structure and its hyperparameters are easy to adjust; i Use decision trees to learn target values y i and i -1 residuals from the sum of the outputs of the decision trees; using a boosting learning strategy, in generating the... k When planting trees, we need to minimize the objective function of the following expression. obj To adjust the first k Tree learning parameters θ ; When the objective function is minimized, the output value is... f k close to y k The final output can be obtained by summing the output values of all decision trees. To achieve the future N The health status synchronous prediction of the step adopts a multi-output forest regressor strategy, by establishing N Each XGBoost model outputs a single SOH step, thereby achieving the goal of multiple outputs.