基于CNN-Bi LSTM的铁氧体磁芯损耗精确模型和小样本迁移学习预测方法

Accurate Ferrite Core Loss Model Based on CNN-BiLSTM and Few-shot Transfer Learning Prediction Method

传统的经验公式和损耗分离公式难以准确计算宽频宽磁密、宽温度范围以及复杂波形激励下铁氧体磁芯的损耗。考虑到磁芯损耗与磁通密度波形的局部和长期特征均相关,基于普林斯顿大学研究者构建的MagNet数据集,采用CNN-BiLSTM建立大样本磁芯损耗预训练模型,损耗预测的平均误差均小于3%,95%误差均小于10%;以3C90和N87铁氧体为例,构建小样本数据集并采用迁移学习方法训练模型,选取最优迁移学习策略,提出最优源域模型选取方法,对比迁移学习和直接训练所需的训练步数,分析小样本数量和初始学习率对迁移学习效果的影响。以样本数量达到1000为例,与直接训练相比,采用迁移学习方法后模型所需训练步数由500降为50,3C90和N87铁氧体损耗预测的平均误差分别由4.49%和6.6%降为2.66%和2.35%,95%误差分别由11.97%和17.12%降为7.22%和6.21%,模型的收敛速度和预测精度都大大提高。在实际工程中,仅需利用少量样本对源域模型参数进行微调,即可实现模型快速求解和损耗精确预测。

Traditional empirical formulas and loss separation formulas cannot accurately calculate ferrite core loss under wide frequency,wide flux density,wide temperature range,and complex waveform excitations.Considering the dependence of core loss on both the local and long-term characteristics of flux density waveform and utilizing the MagNet dataset built by researchers in Princeton University,we established a large-sample core loss pre-training model based on CNN-BiLSTM.The average prediction errors of core losses are all below 3%and the 95%errors are all below 10%.The 3C90 and N87 are taken as examples,few-shot core loss datasets are established,and transfer learning method is applied to train the model.The optimal transfer learning strategy is selected and optimal source model selection method are proposed.The required training steps of transfer learning and direct training are compared.The impacts of few-shot data size and initial learning rate on the transfer learning effect are analyzed.A sample size of 1000 is taken as an example and compared with direct training,the required training steps are reduced from 500 to 50 by adopting transfer learning.The average prediction errors of 3C90 and N87 ferrite core losses are reduced from 4.49%and 6.6%to 2.66%and 2.35%respectively.The 95 percentile prediction errors are reduced from 11.97%and 17.12%to 7.22%and 6.21%,respectively.Both the convergence speed and prediction accuracy of the model are improved.In practical engineering,only few-shot dataset is required to fine-tune the parameters in the source domain model to realize fast model solving and accurate core loss prediction.