Towards Self Driving Processes: Leveraging the Data Revolution

人们普遍认为，我们正处于第四次工业革命的黎明，这将带来前所未有的自动化过程工业水平。这种乐观情绪是由无处不在的网络物理系统相当偶然的融合、轻松访问大量数据、不断扩大的计算能力以及数据分析方面的重大理论突破所激发的。我们集团的长期愿景是创造类似于自动驾驶汽车的自动驾驶过程。通过生成必要的过程输入，通过学习其动态，通过自动调整其控制器，以及通过检测、隔离和预测各种故障，真正的自动驾驶自主过程将在最少甚至没有人工干预的情况下运行。这项提议寻求开发一套算法和计算工具，将自动化的愿景带给流程工业。 工业过程具有复杂的非线性和随机动态、多速率噪声测量和大单元互联等特点。该提案针对每一个过程特征，为自动驾驶过程建立了一个全面的系统方法。为了追求我们的愿景，这个项目分为三个子项目：1)模型监控和主动数据生成：我们将开发算法来主动监控模型的性能过程。这将在不注入外部输入的情况下完成，从而防止性能下降。一旦模型被确定为差的，自动算法将生成一组新的足够信息的数据以用于模型重新识别。2)自适应建模和控制：使用生成的数据，将在线识别新模型并将其嵌入控制策略。根据可获得的信息，这种方法将采取两种形式。当模型结构可用时，将结合仿真方法和非线性模型预测控制策略使用最大似然方法。当模型结构不可用时，将使用深度强化学习等方法来设计控制器。3)故障检测与隔离：在大规模数据集上使用基于仿真的方法和机器学习算法来识别过程故障。特别是，已知故障将通过使用基于模型的算法来识别，未知故障将通过从大维数据集中学习和提取特征来识别。这些算法将在我们的工业合作伙伴的帮助下在实际流程中进行测试。 这一建议是新颖的，因为它对具有相当一般特征的过程采用了独特的统一方法。据我们所知，这是有史以来第一次尝试建立具有上述功能的自动驾驶过程。实现这一愿景将使加拿大工业变得高效，并使它们获得全球竞争优势，从而使加拿大工业受益。

It is widely believed that we are at the dawn of the fourth industrial revolution that will bring a level of automation process industry has never seen before. This optimism is spurred by the rather serendipitous confluence of ubiquitous cyber-physical systems, easy access to large volumes of data, expanding computing power and major theoretical breakthroughs in data analytics. Our group's long term vision is to create self-driving processes similar to self driving cars. A truly self-driving autonomous process will operate with minimal to no human interference by generating necessary process inputs, by learning its dynamics, by automatically tuning its controller, and by detecting, isolating and predicting various faults. This proposal seeks to develop a set of algorithms and computational tools to bring this vision of automation to the process industry. Industrial processes are characterized by complex nonlinear and stochastic dynamics, multi-rate noisy measurements and large interconnected units. This proposal addresses each of these process characteristics to build an over arching systematic approach for self driving processes. In pursuit of our vision, this project is divided into three sub-projects: 1) Model Monitoring and Active Data Generation: We will develop algorithms to actively monitor the performance of a model of the process. This will be done without injecting external inputs and therefore preventing performance degradation. Once a model is determined to be poor an automatic algorithm will generate a new set of sufficiently informative data for model re-identificaiton. 2) Adaptive Modelling and Control: Using the generated data, new models will be identified online and embedded in a control strategy. This approach will take two forms depending on the available information. When a model structure is available a maximum likelihood approach in conjunction with simulation methods and a nonlinear model predictive control strategy will be used. When a model structure is unavailable approaches such as deep reinforcement learning will be used to design the controller. 3) Fault Detection and Isolation: Process faults will be identified using simulation based methods and machine learning algorithms on large scale data sets. In particular known faults will be identified by using model based algorithms and unknown faults will be identified by learning and extracting features from large dimensional data sets. These algorithms will be tested on real processes with the help of our industrial partners. This proposal is novel due to its unique unifying approach for processes with rather generic characteristic features. To the best of our knowledge, this is the first ever attempt to build a self driving process with the features described above. Realizing this vision will benefit Canadian industries by making them highly efficient and by giving them a global competitive advantage.