The performance of off-line build soft sensor often deteriorates,where updating the model online is necessary. For this reason,an approach based on selective kernel learning for on line soft sensing is proposed. This method,utilizing the least squares support vector machine(LSSVM)for constructing offline model,employs the strategy of prediction error bound(PEB)to carry out forward learning selectively so as to enhance model sparsity. Moreover,in order to delete redundant samples more accurate?ly in backward learning,this paper proposes a similarity criterion in high dimensional feature space which incorporates the input and output information simultaneously so that the most dissimilar sample to the current state is selected and eliminated. The forego?ing scheme is applied to build the soft sensor of melt index of polypropylene and the result has demonstrated the effectiveness of the proposed method.%离线建立的软测量模型的预报精度往往逐渐降低,通常需要在线更新模型.为此,提出一种基于选择性核学习的在线软测量建模方法.该方法首先建立最小二乘支持向量机模型,并采用基于预报误差限的选择性前向学习策略更新模型,以提高其稀疏性.此外,为了更准确地删除冗余样本,在后向学习中,提出一种在高维特征空间同时利用输入输出数据的相似度准则,选择性删除与当前状态相似度最小的样本.利用某石化公司聚丙烯熔融指数的软测量建模结果进行了方法验证.
展开▼
