Execution Time Measurement of Narma-L2 Neural Controller Training on Embedded Devices for Mobile Machines

Many machines in the construction domain are driven by hydraulic actuators. The control of these hydraulic systems is highly challenging due to multiple nonlinearities. A promising approach for the optimal control of these hydraulic systems is adaptive control, in which the control parameters are adapted to the changing operating points. For adaptive control, the hydraulic system must be continuously identified. In this paper, an AI-based control approach (Narma-L2 Neural Controller) is presented for the hydraulic system. Based on the requirements of mobile hydraulic-driven machines, some portable heterogeneous hardware platforms are selected to measure and evaluate the execution time of computationally intensive Narma-L2 Neural Controller training. The execution time measurements enable to evaluate and assess which hardware platforms can be selected for adaptive control approaches in mobile machines to avoid overpowered hardware systems. For the demonstration proposed, four embedded single-board computers are chosen that are well established and have a similar behavior in computational power to the controllers often used in mobile systems. The execution time for Narma-L2 Neural Controller training on the most powerful considered device is more than approximately 5 times faster than on the hardware platform with the lowest computational power. The measurements presented in this paper allow to assess whether a low-cost, highly resource-limited hardware is sufficient to execute the algorithms or whether a higher-performance, but also more expensive computer should be used.