一种基于流形正则化随机配置网络的化工过程故障识别方法

A fault identification method of chemical process based on manifold regularized stochastic configuration network

考虑到化工过程故障数据的复杂非线性特性和样本潜在的根本结构特征,提出了一种基于流形正则化随机配置网络的故障识别方法。该方法在经典随机配置网络的基础上,在嵌入流形约束的监督机制下随机选取隐含参数,逐个添加隐含节点,然后使用流形正则化最小二乘法计算隐含层的输出权值,保留了数据的重要几何特征,避免了信息冗余,有利于更好地识别出区别于不同类别的相关特征。在测试集上的实验表明,该方法对TE故障和半导体故障的识别准确率分别达到了87.72%和84.27%,均高于随机向量函数连接网络和随机配置网络方法。而且对于大部分故障类型,该方法的精确率和召回率较高,验证了所提方法进行故障识别的有效性和所建立模型的良好泛化能力。

Considering the complex nonlinear characteristics of chemical process faults and the underlying structural characteristics of samples,a fault identification method based on manifold regularized stochastic configuration network is proposed.Based on classical stochastic configuration network,this method randomly selects hidden parameters under the supervision mechanism of embedded manifold constraints to add hidden nodes one by one.Then,the output of hidden layer weights is calculated by manifold regularized least square method.It keeps the important geometric characteristics of data.The information redundancy is avoided and the relevant characteristics of different from different categories could be identified.Experimental results on test set show that the identification accuracy values of TE fault and semiconductor fault are 87.72%and 84.27%,respectively,which are higher than those of random vector function connection network and stochastic configuration network.In addition,for most fault types,the precision and recall rates of the proposed method are high.Results prove that the proposed method can effectively identify faults.The generalization ability of fault identification model is improved.