Self-adaptive modelling of large-scale process systems using machine learning

过程系统已经可以通过各种方法来实现，这些模型往往最终会变成对真实系统的不准确描述，从而导致工厂的次最优运行。出现这种不匹配的原因有很多，例如底层系统的变化(如结垢和设备变化)。因此，需要训练有素的工程师重新检查问题并维护基本模型。这可能会产生重大费用，对于存在不匹配的操作期间，这也可能导致次优的工艺操作，从而导致重大利润损失。此外，这种对工程工作的依赖为大型加工厂部署基于模型的解决方案带来了可伸缩性问题。如果目标是在行业内达到更高的自治度，所有这些问题都特别相关，因为这些问题是实现自主工业系统目标的障碍。基于上述原因，本博士学位的目标将是开发数据驱动的自适应建模框架，该框架可以处理底层系统的变化并相应地进行调整，以保证过程模型的准确性。如果能够实现这一点，对大规模工业流程的影响可能是巨大的。这可能通过多种方式实现，但本博士学位的主要重点将是利用统计学习技术，范围从简单的机器学习方法到更复杂的深度学习算法。虽然其中许多技术对这项任务显示出很大的前景，但在这些技术对大规模过程系统的动态建模的适用性、可扩展性、培训的稳定性以及实现这些目标所需的数据量和种类等方面仍然存在一些问题。因此，在这篇博士论文中，我们将研究这样的公开问题。

process systems can already be achieved by various means, often these models will eventually become an inaccurate representation of the real system, which results in the sub-optimal operation of the plant. This mismatch can begin to occur due to a number of reasons such as changes in the underlying system (e.g. fouling and equipment changes). Consequently, there is a need for highly-trained engineers to re-visit the problem and maintain the underlying models. This can incur major expenses and for the period of operation where the mismatch existed, this can also result in sub-optimal process operation resulting in the loss of significant profits. Additionally, this dependence on engineering effort creates a scaling problem for deploying model-based solutions for large process plants. All of these issues are especially relevant if the objective is to reach a higher degree of autonomy within industry as these issues act as obstacles to achieve the goal of autonomous industrial systems.For the above reasons, the objective of this PhD will be to develop data-driven self-adaptive modelling frameworks which can handle changes in the underlying system and adapt accordingly to guarantee the accuracy of the process model. If this can be achieved, the impact on large-scale industrial processes could be significant. This could potentially be achieved in a number of ways, but the main focus of this PhD will be to utilize statistical learning techniques ranging from simple machine learning methods to more complex deep learning algorithms. While many of these techniques have shown great promise for the task, some questions still remain regarding the applicability of these techniques to dynamic modelling of large-scale process systems, their scalability, the stability of training and the amount and variety of data required to achieve these objectives. Consequently, in this PhD we will investigate such open problems.