非平稳间歇过程的监测与智能自愈控制研究

Characterized by fast product switch, small amount of production, high-value added product, batch processes have increasingly attracted lots of attentions. Particularly, safe operation and high-quality products are indispensable for batch manufacturing processes. By deeply understanding the operation mechanism, batch processes experience multi-stage operation, large-scale non-stationarity, accumulative quality etc. Moreover, non-stationarity is the inherent mechanism of batch processes. In this project, high efficient operation is our main purpose. With the non-stationarity analysis, a set of complete methodologies have been well established, including process monitoring, quality prediction, and self-healing control, based on data-driven and intelligent methods. It is the first time that monitoring and identification issues of startup stage are analyzed. In this way, it is able to indicate when the production process enters into stable production stage. Besides, the relationship between non-stationarity and accumulative characteristic has been studied and a global prediction model is developed for stable production stage. Finally, to deal with non-optimal conditions, a self-healing adjustment strategy has been developed based on non-optimality operation rules and reinforcement learning.

具有快速产品切换、小批量生产、高附加值等诸多优点，间歇过程受到持续广泛的关注，其安全可靠运行以及高品质产品质量成为我国智能制造不可或缺的内容。通过深入挖掘其运行机理，发现其具有多阶段运行、大跨度非平稳、累积质量等特性，特别是非平稳特性是间歇过程的内在本质驱动。本项目以间歇过程高效运行为出发点，首次从非平稳特性分析入手，基于数据挖掘与人工智能方法，系统建立面向过程监测、质量预测以及智能自愈理论的一整套完整体系。该体系首次涵盖启动阶段的研究，建立基于多元统计分析的启动阶段辨识和在线监测方法，达到及时、准确指示生产过程何时进入稳态阶段的目的。其次，进一步揭示累积质量特性是非平稳特性在质量指标上的体现这一内在机理，完善稳态运行阶段下的全局化累积质量预测关系的建立。最终，构建非优运行状态判断准则和建立基于强化学习的自愈调控机制，实现过程的高效运行。

间歇模式的安全可靠运行以及高品质性能成为我国智能制造不可或缺的部分。通过深入挖掘其运行机理，发现其具有多阶段运行、大跨度非平稳等特性，特别是非平稳特性是间歇过程的内在本质驱动。本项目以间歇模式高效运行为出发点，将新一代人工智能技术与工业过程建模与控制深度融合，开展了动态环境下数据驱动的建模和控制研究，实现了间歇模式的自主决策与自愈控制。首先，针对间歇模式启动过程的非平稳特性，提取不随启动过程快速动态特性发生变化的平稳分量，建立了系统性的启动阶段监测和划分模型。其次，针对建模数据和实际应用数据之间存在的差异，分解出时序特征用于构建基于长短记忆网络的基准预测模型；进一步，基于多变量统计分析原理建立面向差异性变化分析的时序特征监测模型，及时在差异性过大而导致预测精度严重下降时发出预警并建立快速的迁移策略。最后，针对过程中发生的故障或异常对最优运行状态造成的影响进行深入分析，建立了一种基于面向不确定的强化学习方法。通过利用贝叶斯卷积神经网络对性能和不确定性进行建模聚合，并利用基于Q-Learning的强化学习算法用于设定点决策。..相关研究成果发表在工业信息和自动化领域顶级期刊《IEEE Transactions on Information Informatics》、《IEEE Transactions on Cybernetics》、《Applied Energy》、《Chemical Engineering Research and Design》等相关领域期刊和会议发表SCI/EI论文9篇。研究成果丰富和深化了复杂工业过程的安全运行理论，推动了基于模型和知识的先进控制技术向基于数据的智能控制技术的发展。