基于特征子空间虚假邻点判别的铝电解槽况诊断模型

Diagnosis Model of Status of Aluminum Reduction Cells Based on False Nearest Neighbors in Feature Subspace

造成铝电解异常槽况的因素较多,彼此相关性强,建立槽况诊断模型时计算量巨大。利用核主元分析法虽然可以对非线性数据进行降维,但得到的主元仍然是原始变量在特征空间的线性组合,既无明确的物理含义,又无法对原始特征进行约简达到减少传感器种类的目的。提出一种基于特征子空间虚假邻点判别的槽况诊断方法,首先考察各原始变量置零前后在核空间主元投影上的相似度,根据其对槽况的解释能力进行原始变量选择;再将约简后的原始变量输入概率神经网络,对各类异常槽况进行诊断。通过取自某厂170KA大型预焙槽的268组样本进行检验:在原始特征约简44.4%的情况下分类精度达到95%以上,表明该方法不但可对原始特征进行有效约简,而且槽况分类精度和训练时间均优于同类模型。

There is huge computation when status diagnosis model for aluminum electrolytic cells is established because of many strong correlation factors.Kernel principal component analysis can be used to reduce the dimensionality of the nonlinear cell data;however,the principal components haven＇t clear physical meaning as a linear combination of the original variables in the feature space.The method can not used to reduce original feature in order to achieve the purpose of sensors reduction.To overcome the above-mentioned problems,a novel diagnosis method based on false nearest neighbors（FNN） in feature subspace is proposed.In the proposed approach,it is inspired by FNN that interpretation of alumina concentration would be estimated by calculating the variables mapping distance in the kernel principal components analysis（KPCA） feature subspace.Selected variables are introduced into probabilistic neural network（PNN） as input vector to diagnose and classify different status of aluminum reduction cells.By using 268 groups of sample of 170KA operating aluminum cell from a factory,experimental results demonstrate that the classification accuracy is 95％ and the original feature is reduced to 44.4％.The results show that the original feature is reduced effectively and classification accuracy and training time of diagnose five status of aluminum reduction cells are better than similar models.