基于KPCA和数据处理组合方法神经网络的半球谐振陀螺温度建模补偿方法

Temperature modeling and compensation method of hemispherical resonator gyro based on KPCA and grouped method of data handling neural network

针对半球谐振陀螺(HRG)的温度建模与补偿问题,提出基于核主成分分析(KPCA)和数据处理组合方法(GMDH)神经网络的建模补偿方法.通过分析HRG的温度特性和大数据特征,初步确定网络模型的特征向量.为了去除HRG输出数据的相关性和冗余性,引入KPCA并降低特征向量维度.将特征向量代入GMDH神经网络训练,区分训练集和验证集以确定网络权值和网络结构,实现HRG温度漂移的建模与补偿.实验结果表明,单一样本预测时,所提方法预测效果明显好于传统多项式模型;多样本预测时,在4种不同训练样本下,所提方法相比传统多项式模型精度分别提升了48.5%、54.0%、56.3%、68.4%,相比GMDH模型分别提升了3.6%、5.1%、3.8%、8.8%.所提方法能够有效提高HRG在变温工况下的测量精度.

A modeling and compensation method based on kernel principal component analysis(KPCA)and grouped method of data handling(GMDH)neural network was proposed aiming at the temperature modeling and compensation of hemispherical resonator gyro(HRG).By analyzing the temperature characteristics and the big data characteristics of HRG,the eigenvectors of the network were initially selected.To remove the correlation and redundancy of the HRG outputs,KPCA was introduced and the eigenvector dimension was reduced.The eigenvectors were substituted into the GMDH neural network and the training set and the validation set were distinguished to determine the network weight and structure to model and compensate for the HRG temperature drift.Experiment results showed that the proposed method was significantly better than the traditional polynomial model for single-sample predictions;for multiple-sample predictions,under four different training samples,the accuracy of the proposed method was 46.5%,51.5%,54.6%and 65.3%higher than that of the traditional polynomial model,also 3.6%,5.1%,3.8%and 8.8%higher than that of the GMDH model.The proposed method effectively improved the measurement accuracy of HRG under variable temperature conditions.