IPMC率相关迟滞特性迁移学习建模  

The transfer learning modeling of IPMC rate-dependent hysteresis

作　　者：白亭亭 孟江[1] 王格 曹凤蓉 安坤 Bai Tingting;Meng Jiang;Wang Ge;Cao Fengrong;An Kun(School of Mechanical Engineering,North University of China,Shanxi Taiyuan,030051,China;School of Electrical and Control Engineering,North University of China,Shanxi Taiyuan,030051,China)

机构地区：[1]中北大学机械工程学院,山西太原030051 [2]中北大学电气与控制工程学院,山西太原030051

出　　处：《机械设计与制造工程》2025年第3期116-121,共6页Machine Design and Manufacturing Engineering

基　　金：山西省面上自然科学基金(202203021221098)。

摘　　要：针对离子聚合物金属复合物(IPMC)致动器在低频交流驱动电压下具有明显的率相关迟滞非线性,提出一种卷积神经网络(CNN)结合基于模型迁移学习的率相关迟滞建模方法。首先搭建实验平台,在0.1~3.0 Hz、2.0~5.5 V线性递增正弦驱动电压下,测得电压-位移实验数据,构建CNN模型对足量源域数据进行预测,利用粒子群优化算法对CNN模型关键参数进行寻优,得到预训练模型;然后采用基于模型的迁移学习方法,利用少量目标域数据,对预训练模型中的部分网络参数进行微调;最终获得可以对不同频率目标域数据有着良好预测能力的新模型,从而实现在目标域数据稀少情况下的率相关迟滞建模。通过该模型与LSTM、GRU模型进行对比验证,结果显示CNN迁移学习模型在不同频率下迁移精度均在98.23%以上,平均训练时间缩短46.1%。Aiming at the obvious rate-dependent nonlinear hysteresis of IPMC actuators under low frequency AC drive voltage,a transfer learning modeling method based on the convolutional neural network(CNN)model transfer learning is proposed for the rate-dependent hysteresis.Firstly,an experimental platform is built,and the volt-displacement experimental data are measured under the increasing sinusoidal driving voltages varying 0.1~3.0 Hz and 2.0~5.5 V,the CNN model is constructed to predict the sufficient source data,and the key parameters of CNN model are optimized by the particle swarm optimization algorithm to obtain the pre-trained model.Secondly,a model-based transfer learning method is used to fine tune certain network parameters of the pre-trained model with a small amount of target domain data.Finally,a new model is obtained to achieve good prediction ability for different frequency target domain data,so as to achieve rate-dependent hysteresis modeling in the case of sparse target domain data.Compared with LSTM and GRU models,the results show that the transfer accuracy of CNN transfer learning model is more than 98.23%at different frequencies,and the average training time is reduced by 46.1%.

关 键 词：率相关迟滞特性 卷积神经网络 粒子群优化算法 迁移学习 微调 

分 类 号：TB381[一般工业技术—材料科学与工程]