基于Preisach模型的深度学习网络迟滞建模

Hysteresis modeling with deep learning network based on Preisach model

中文关键词  压电扫描器  迟滞非线性  Preisach模型  深度学习

英文关键词  piezoelectric scanner  hysteresis nonlinear  Preisach model  deep learning

中文摘要

      针对传统压电扫描器迟滞模型泛化能力较弱的问题, 提出了一种基于Preisach 模型的深度学习网络来建立迟滞模型, 提高了模型的学习能力和泛化能力. 具体而言, 首先利用深度学习在深度特征提取方面的优势, 建立包含卷积层、池化层、展开层以及深度特征层的深度学习层来提取输入电压信号的特征信息; 其次, 利用傅里叶变换层计算得到输入信号的频率, 并将频率输入到非线性层, 构造并输出了与输入信号频率相关的非线性项, 该非线性项作为权值函数与Preisach模型的迟滞单元输出相乘, 并将乘积叠加得到了频率相关的模型输出向量; 最后, 将深度学习层输出的特征向量与Preisach模型输出向量点乘, 即可得到深度学习网络的最终输出位移. 同时利用电容位移传感器采集的16 组输入输出信号对深度学习网络进行训练, 得到了网络中的权值参数, 并利用其他8 组输入输出数据对深度网络进行测试, 训练和测试结果表明, 本文所提出的基于Preisach 模型的深度学习网络在得到高精度迟滞模型的同时, 提高了模型的泛化能力.

英文摘要

      Aiming at the weak generalization ability of traditional piezoelectric scanner hysteresis models, a deep learning network based on Preisach model is proposed to establish the hysteresis model for piezoelectric scanners, which improves the learning and generalization ability of the model. Specifically, first, considering the advantage of deep learning network in feature extraction, a deep learning layer comprising two convolution layers, a pool layer, an expansion layer, and a deep feature layer is established to extract the characteristic information of the input voltage signal. Afterwards, a Fourier transform layer is used to calculate the frequency of input signal, which is then input to the nonlinear layer to output a frequency-dependent nonlinear term, subsequently, the nonlinear term is multiplied by the hysteresis unit of the Preisach model to obtain the frequency-dependent model output vector. Finally, the output displacement of the whole depth learning network is obtained by multiplying the feature vector of the depth learning layer with the output vector of the Preisach model. In the section of network training and testing, 16 groups of input and output signals collected by the capacitance displacement sensor are used to train the deep learning network to get the weight parameters, and the other 8 groups of input and output data are tested on the deep network. The results show that the proposed deep learning network improves the generalization ability of the model while obtaining the high precision hysteresis model.