Scholarlink

<p>The real-time implementation of nonlinear model predictive control (NMPC) is challenging due to the computational complexity of solving online optimization problems within a stipulated sample time. A popular solution to this issue involves approximating NMPC control laws using supervised learning with deep neural networks (DNNs). The resulting DNN controller offers faster execution and has a smaller memory footprint than the standard NMPC methods. However, guaranteeing constraint satisfaction remains crucial in this imitation learning framework. To address this challenge, one approach is to incorporate constraint information during the neural network training phase. This study implements the penalty and logarithmic barrier methods to enrich the neural network’s loss function with insights regarding constraints. Using these methods, the neural network can handle constraints, ensuring adherence to them during control decisions. To validate the proposed approach’s efficacy, it is evaluated using a 2-degree-of-freedom (DOF) nonlinear helicopter model through deployment on STM32 Nucleo-144 development board using hardware-in-the-loop (HIL) co-simulation. The performance of the proposed controller is compared to that of the standard NMPC. The comparative analysis demonstrates that the proposed DNN algorithm not only reduces memory consumption and controller execution time but also effectively handles the system constraints, delivering performance comparable to standard NMPC.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Embedded DNN-based nonlinear MPC with guaranteed constraint satisfaction

  • 期刊论文

  • 发表日期：

    2025年3月29日

查看全文
  • Nirlipta Ranjan Mohanty,
  • Promod Ubare,
  • Vaishali Patne,
  • Deepak Ingole,
  • Dayaram Sonawane

摘要

The real-time implementation of nonlinear model predictive control (NMPC) is challenging due to the computational complexity of solving online optimization problems within a stipulated sample time. A popular solution to this issue involves approximating NMPC control laws using supervised learning with deep neural networks (DNNs). The resulting DNN controller offers faster execution and has a smaller memory footprint than the standard NMPC methods. However, guaranteeing constraint satisfaction remains crucial in this imitation learning framework. To address this challenge, one approach is to incorporate constraint information during the neural network training phase. This study implements the penalty and logarithmic barrier methods to enrich the neural network’s loss function with insights regarding constraints. Using these methods, the neural network can handle constraints, ensuring adherence to them during control decisions. To validate the proposed approach’s efficacy, it is evaluated using a 2-degree-of-freedom (DOF) nonlinear helicopter model through deployment on STM32 Nucleo-144 development board using hardware-in-the-loop (HIL) co-simulation. The performance of the proposed controller is compared to that of the standard NMPC. The comparative analysis demonstrates that the proposed DNN algorithm not only reduces memory consumption and controller execution time but also effectively handles the system constraints, delivering performance comparable to standard NMPC.