Model-based control is preferred for robotics applications due to its systematic approach to linearize and control the robot's nonlinear dynamics. The fundamental challenge involved in implementing a model-based controller for robotics applications is the time delay associated with the real-time computation of the robot dynamics. Due to the sequential structure of the robot's dynamic equation of motion, the multicore CPU cannot reduce the control algorithm execution time. A high-speed processor is required to maintain a higher sampling rate. Neural network-based modeling offers an excellent solution for developing a parallel structured equivalent model of the sequential model that is suitable for parallel processing. In this paper, a Deep neural network-based parallel structured 7 degrees of freedom human lower extremity exoskeleton robot controller is developed. Forty-nine densely connected neurons are arranged in four layers to estimate joint torque requirements for tracking trajectories. For training, the deep neural network, an analytical model-based data generation technique is presented. A trained deep neural network is used for real-time joint torque prediction and a PD controller is incorporated to mitigate the prediction errors. Simulation results show high trajectory tracking performances. The developed controller's stability analysis is proved. The robustness of the controller against the parameter variation is analyzed with the help of the analysis of variance (ANOVA). A comparative study between the developed controller and the Computed Torque Controller, Model Reference Computed Torque Controller, Sliding Mode Controller, Adaptive controller, and Linear Quadratic Regulator are presented while keeping the same robot dynamics.

翻译：基于模型的控制因其能够系统化地线性化并控制机器人非线性动力学特性，在机器人应用中备受青睐。但在机器人应用中实现基于模型的控制器的核心挑战，在于实时计算机器人动力学所带来的时间延迟。由于机器人运动方程具有顺序结构，多核CPU无法降低控制算法的执行时间，因此需要高速处理器来维持较高的采样率。基于神经网络的建模为开发适用于并行处理的顺序模型并行结构等效模型提供了绝佳解决方案。本文开发了一种基于深度神经网络的并行结构七自由度人体下肢外骨骼机器人控制器。其通过四层共四十九个密集连接的神经元来估计跟踪轨迹所需的关节扭矩。为训练深度神经网络，本文提出了一种基于分析模型的数据生成技术。训练后的深度神经网络用于实时预测关节扭矩，并引入PD控制器来减小预测误差。仿真结果表明其具有较高的轨迹跟踪性能，并验证了所开发控制器的稳定性。通过方差分析（ANOVA）评估了该控制器对参数变化的鲁棒性。在保持相同机器人动力学特性的前提下，将所开发的控制器与计算扭矩控制器、模型参考计算扭矩控制器、滑模控制器、自适应控制器及线性二次型调节器进行了对比研究。