基于进化多目标优化的集成即时学习软测量建模方法研究

Soft sensor technology has become a promising solution towards online real-time estimations of difficult-to-measure variables in the process industry, which is of great importance for facilitating efficient monitoring, controlling and optimization of industrial processes. However, many of modern industrial processes are characterized by inherent nonlinearity, time-varying behavior and multiplicity of operation modes /phases that pose significant challenges to develop high-performance soft sensors. For such complex processes, conventional soft sensors often encounter poor prediction accuracy and adaptability since they can only handle the minority of process characteristics, which has become the bottleneck for promotion and application of soft sensor technology. Therefore, the project, beginning with traditional just-in-time learning (JIT) based soft sensors, focuses on research of ensemble just-in-time learning (EJIT) based soft sensor modeling methods by introducing ensemble learning. The main contents are summarized as follows. Firstly, the multimodal perturbation mechanism of generating diversity of JIT learning models is investigated. Then the building of multiple diverse and accurate JIT learning models is transformed into a multi-objective optimization problem and is further solved by using evolutional multi-objective optimization techniques. Furthermore, the online dynamic combination of diverse JIT learning models is achieved by using ensemble learning. Finally, the adaptation mechanism of EJIT learning based soft sensors is investigated. Overall, the project aims to establish a unified soft sensor modeling framework using EJIT learning, and thus provides new insights into the development of high-performance soft sensors. The resulting achievements will contribute to the promotion and application of soft sensor technology, as well as greatly improve the product quality and control performance of typical industrial processes such as chemical, metallurgy and pharmaceutical production processes.

软测量技术是实现难测参数在线检测的一种有效途径，对工业生产过程的监测、控制及优化具有重要意义。然而，很多工业过程往往呈现出非线性、时变性、多模式、多时段等复杂特性，而常规的软测量方法仅限于处理少数过程特征，从而导致模型预测性能不佳、适应性不强，这成为制约软测量技术推广应用的主要瓶颈。因此，本项目以传统即时学习方法为出发点，通过引入集成学习思想，开展集成即时学习软测量建模方法的研究。具体包括：研究多模态多样性扰动机制，将多样性即时学习模型的构建转为一个多目标优化问题，并基于进化多目标优化技术实现问题求解；研究多样性即时学习模型的在线动态融合策略；研究集成即时学习软测量模型的自适应更新机理。本项目旨在建立一个统一的集成即时学习软测量建模框架，为建立高性能的软测量模型提供新的思路。项目的研究成果有助于促进软测量技术的推广应用，对提高化工、冶金、制药等典型生产过程的产品质量和控制性能具有重要意义。

项目以即时学习为基础，结合集成学习和进化优化技术，提出了一套用于复杂工业过程的数据驱动软测量建模方法。项目实施四年来，主要取得如下成果：.针对常规即时学习中单一相似度性能受限的问题，定义了一种混合加权相似度准则，并采用混合整数遗传算法实现参数优化；针对传统的启发式多样性基模型生成方法难以有效确保基模型多样性和准确性，提出了一种基于进化多目标优化的多样性即使学习基模型生成方法；针对集成建模基模型选择困难的问题，提出了一种基于进化多目标优化的集成修剪策略；针对标记样本不足导致监督模型性能受限的问题，提出了一种基于进化优化的伪标记样本估计方法。. 针对传统即时学习软测量算法仅使用单一学习设置导致无法有效处理复杂过程特征的问题，通过融合基于进化多目标优化的多样性即时学习模型生成方法和集成学习策略，提出了基于多样性加权相似度、多样性异构相似度的集成即时学习软测量建模方法。. 针对基于单一扰动机制构建的集成学习和集成即时学习模型在预测性能上存在的局限性，提出了融合非标记样本扰动和模型参数扰动的选择性半监督集成软测量方法；基于深度学习多样性特征提取的选择性集成方法；基于相似度扰动和输入特征扰动的集成即时学习方法；基于相似度扰动、基模型结构扰动和输入特征扰动三重多样性扰动机制的选择性集成即时学习方法；基于相似度扰动和非标记样本扰动的半监督集成即时学习软测量建模方法。.针对工业过程中非线性、多时段、多模式、时变性复杂过程特性容易导致模型性能恶化的问题，通过融入局部学习、在线选择性集成、有限混合自适应集成机制，提出了基于多相似度局部状态辨识的自适应集成学习软测量建模方法。. 通过数值仿真、青霉素发酵过程仿真、TE化工过程、脱丁烷塔过程和工业金霉素发酵过程等案例验证了上述成果的有效性和优越性，为解决复杂过程对象的难测参数预测提供了一种有效的技术手段。