Abstract: Process industry are usually characterized by complex process characteristics such as nonlinearity,multiplicity of phases and modes,and time-varying behavior,which leads to poor prediction performance for traditional global and ensemble soft sensors.Thus,an adaptive soft sensor modeling method named Multi-Similarity based Online Selective Ensemble (MSOSE) was proposed based on ensemble learning with multi-similarity local state identification.Its implementation included three main stages.In the offline modeling stage,the local process states were identified by using different similarity criteria,and then a set of diverse local models was built.In the online prediction stage,the online dynamic selection of local models,the adaptive determination of model weights and the fusion of local prediction results were achieved through a two-level ensemble strategy.In the update phase,Kullback-Leibler (KL) divergence was used to evaluate the difference between the current and the adjacent state data distributions to achieve real-time detection of concept drift,and then decide whether to add a local model online or not.Moreover,the obtained offline analysis data were added to the modeling database.The effectiveness and superiority of MSOSE were verified through an industrial chlortetracycline fermentation process and an industrial debutanizer process.

Key words: soft sensor, ensemble learning, adaptive, time-varying behavior, multi-similarity, local state identification, Gaussian process regression

摘要： 鉴于流程工业过程的非线性、多时段、多模式、时变性等复杂过程特性,导致传统的全局和集成学习软测量方法预测性能不佳,提出一种基于多相似度局部状态辨识的集成学习自适应软测量建模方法。该方法在离线建模阶段,从不同相似度准则出发辨识局部过程状态,进而生成多样性的局部模型;在线预测阶段,通过双层集成策略实现模型的在线动态选择、模型权重自适应确定、局部预测结果融合;在线更新阶段,通过KL散度评价当前与相邻状态数据分布的差异性以实现概念漂移的实时检测,进而决定是否在线添加局部模型。将新获取的离线检测值添加入建模数据库。通过工业金霉素发酵过程和脱丁烷塔过程验证了所提方法的有效性和优越性。

关键词: 软测量, 集成学习, 自适应, 时变特征, 多相似度, 局部状态辨识, 高斯过程回归