Next Generation Data Driven Modeling and Control of Batch and Batch Like Processes

有许多使用批处理的产品，如药品和特种化学品。几乎每一个进程的启动都涉及一个类似批量操作的过程，将进程从停机模式带入连续运行。目前，这些过程中的许多都使用配方来生成符合规格的产品(产品质量通常只有在批次终止时才能测量)，这有两个主要缺陷。首先，这些配方的开发非常昂贵和耗时(例如，想想为新药开发生产配方所需的时间)，第二，当原材料发生变化时，这些配方不能很好地发挥作用。本研究建议从现有配方作业中收集数据，建立工艺配方与产品质量之间的模型，进而为新产品快速开发配方，并创建在线控制算法来维护按规格的产品。该计划将利用PI小组在数据驱动建模方面的进步来创建工具，这些工具结合了基于机器学习的方法(仅在适当的情况下)，以及较新的传感技术(如图像和声学)的可用性。生物制药和旋转成型的应用将被用来开发方法和展示概念验证。这些计划目标的影响将是迅速的-使用旋转成型装置(目标是生产回收塑料产品)使用线性模型进行快速产品开发的第一线将导致旋转成型行业(例如，Resraft Inc.)的直接利用。以降低生产成本和直接对环境产生积极影响的方式。神经网络和子空间识别方法的利用将为这些技术的使用铺平道路，例如在生物反应器的情况下。众所周知，生物反应器具有非线性和复杂的动力学，使用这些技术将使生物制药的生产成本降低(通过由Sartorius Inc.等工业合作伙伴利用)。图像等非传统数据的使用将影响广泛的行业，如钢铁行业，在该行业，高温使传统传感器的使用变得困难(以及旋转成型和生物反应器)。更重要的是，由于将开发的工具将是一般性的，它们不仅适用于各种工业合作伙伴(作为麦克马斯特先进控制联盟的一部分)，而且适用于加拿大各地的几家制造商。预计到该计划的第四年，短期目标的好处将开始显现，到五年计划结束时，将为加拿大制造业每年节省数十万美元的开发和运营成本，并为实现创造自动快速产品设计和控制工具的长期目标做好准备。

There are numerous products created using batch processing, such as pharmaceuticals and specialty chemicals. The startup of almost every process involves a batch like  operation that takes the process from shutdown mode to continuous operation. Many of these processes currently employ recipes to generate on spec product (with the product quality typically measured only at batch termination)- and this has two major drawback. First - the development of these recipes is very expensive and time consuming (think e.g, of the time required in developing production recipes for new medication), and the second is that these recipes do not work well when the raw material changes. The present research proposes to collect data from existing recipe based operation, create a model between process `recipe' and product quality to in-turn develop recipes rapidly for new products, and to create online control algorithm to maintain on-spec products. The program will leverage advances in data driven modeling from the PI's group to create tools that incorporate machine learning based approaches (only where appropriate), and availability of newer sensing technologies (such as images and acoustics). Applications to biopharmaceuticals and rotational molding will be used to both develop the approaches and demonstrate proof of concept. The impact of these program goals will be quick- the first thread on rapid product development using linear models using a rotational molding setup (with a goal to produce, for instance, recycled plastic products) will lead to a direct utilization by the rotomolding industry (e.g., Rescraft Inc.) in ways that will cut down production cost and directly impact the environment positively. The utilization of neural networks along with subspace identification methods will pave the way for use of these techniques, for instance, in the case of bioreactors. Bioreactors are known to have  nonlinear and complex dynamics, and using these techniques will make production of biopharamaceuticals less expensive (through utilization by industrial partners such as Sartorius Inc). The use of non-traditional data such as images will impact a huge range of industries, such as the steel industry, where high temperatures make use of traditional sensors difficult (along with rotational molding and bioreactors). More importantly, since the tools that will be developed will be of a general nature, they will be readily applicable not just to various industrial partners  (as part of the McMaster Advanced Control Consortium) but to several manufactureres all over Canada. It is anticipated that by year 4 of the program, the benefits of the short term goals will already start manifesting and by the end of the five year program, will save Canadian manufacturing to the tune of hundreds of thousands of dollars a year in development and operational cost, and be well set for accomplishing the longer term goal of creating an auotomated rapid product design and control tool.