Deep Learning Enabled Maintenance Free Control

深度学习支持免维护控制

• 批准号: 536418-2018
• 负责人: Gopaluni, Bhushan
• 金额: $ 8.1万
• 依托单位: University of British Columbia
• 依托单位国家: 加拿大
• 项目类别: Collaborative Research and Development Grants
• 财政年份: 2020
• 资助国家: 加拿大
• 起止时间: 2020-01-01 至 2021-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=698916
• 关键词: Deep; Learning; Enabled; Maintenance; Free

Modern industrial plants are highly integrated and complex. As such there are several layers of controllers that are commonly used. These include simple single loop PID controllers, multivariable Model Predictive Controllers (MPC), and Real Time Optimizers. These controllers rely on a number of tuning parameters that in turn dictate the overall closed loop performance. Tuning these parameters is an inherently difficult exercise as there are several tuning parameters and they are often not amenable to reliable optimization formulations. In addition, tuning often is resource intensive, expensive and time consuming. Recently, several longstanding problems in computer science have been solved using ideas from dynamic programming and artificial neural networks. The common theme across the solutions developed for these problems is the superior performance of Deep Neural Networks (DNNs) in approximation complex functions and the automatic learning ability of reinforcement learning (RL) algorithms. In this project, we will extended the successes of DNN and RL to automatically tune PID, MPC and other nonlinear controllers.

现代工业工厂高度集成和复杂。因此，通常使用的控制器有几层。这些控制器包括简单的单回路PID控制器、多变量模型预测控制器（MPC）和真实的时间优化器。这些控制器依赖于许多调谐参数，这些参数又决定了整体闭环性能。调整这些参数是一个固有的困难的练习，因为有几个调整参数，他们往往不服从可靠的优化公式。此外，调优通常是资源密集型的、昂贵的和耗时的。 最近，计算机科学中的几个长期存在的问题已经使用动态规划和人工神经网络的思想得到了解决。为这些问题开发的解决方案的共同主题是深度神经网络（DNN）在逼近复杂函数方面的上级性能和强化学习（RL）算法的自动学习能力。在本项目中，我们将扩展DNN和RL的成功，以自动调整PID、MPC和其他非线性控制器。