Hybrid derivative functions for identification of unknown loads and physical parameters with application on slider-crank mechanism

Groote, Wannes De; Kikken, Edward; Goyal, Srajan; Hoecke, Sofie Van; Hostens, Erik; Crevecoeur, Guillaume

doi:10.1109/aim.2019.8868376

Cited by 3 publications

(2 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…11a illustrates the convergence history of three physical parameters p (related to the rotational inertia J 1 , the rotational damping coefficient B m and the mass m 3 of the slider) in the gradient based optimization process. In [38] we show on numerical simulation data that the obtained values after convergence are indeed the true system parameters. The reason for their convergence lies in the high influence these parameters have on the model behavior, illustrated in the sensitivity analysis in Appendix B.…”

Section: B Simultaneous Nnap Model Optimizationmentioning

confidence: 82%

“…The model η is only precise locally around the area because data-driven models are inherently difficult in extrapolation capabilities. In [38] we show on numerical simulation data that the ReLU network η indeed converges towards the real force relation F at these explored areas. The corresponding side view, depicted in Fig.…”

Section: Retrieving Physical Insightsmentioning

confidence: 85%

See 1 more Smart Citation

Neural Network Augmented Physics Models for Systems with Partially Unknown Dynamics: Application to Slider-Crank Mechanism

De Groote,

Kikken,

Hostens

et al. 2019

Preprint

Self Cite

View full text Add to dashboard Cite

Mechatronic systems are plagued by nonlinearities and contain uncertainties due to, amongst others, interactions with their environment. Models exhibiting accurate multistep predictive capabilities can be valuable in the context of motion control and design of servo controlled systems. Neural Network Augmented Physics (NNAP) models are presented in this paper to comprehend the behavior of servo systems that contain partially unknown dynamics. By means of a hybrid modeling formalism, neural network models are closely merged with physics-based state-space derivative functions. The methodology is applied on an experimental slider-crank system consisting of a servo drive and mechanical links of which the physical behavior is only partially known. Its interaction with the environment due to a compression spring load and friction phenomena, however, remains unknown. Accurate prediction capabilities of the presented NNAP models are demonstrated on the slider-crank system. Compared to the feedforward modeling formalism, recurrent NNAP models had the ability to even further reduce prediction errors up to 33.4% on validation data. From the converged dynamic NNAP model we extracted the neural network and identified the unknown phenomena, being the spring characteristic and the friction forces, within the mechatronic system.

show abstract

Section: B Simultaneous Nnap Model Optimizationmentioning

confidence: 82%

Section: Retrieving Physical Insightsmentioning

confidence: 85%