Deep Learning-Based Wrench Model for Magnetically Levitated Actuators

Magnetic levitation actuators (MLAs) are employed in flexible manufacturing, precision positioning and machining, and haptic devices due to untethered motions. However, the MLAs with disc-magnet movers and static coils rely on multidimensional and memory-expensive lookup tables (LUTs) for operation, lacking online force and torque (wrench) models. We propose a deep neural network model to predict the wrench between individual disc magnets and each coil using a programmable logic controller. The approach utilizes residual blocks to deepen the architecture without enlarging network scales, and we train and validate the wrench model using datasets generated with random mover poses inside the operating range. We show that after training, the residual-based model outperforms a two-hidden-layer baseline model in implementation performance and test accuracy. The test/prediction accuracy is verified through load cell measurements and LUTs on two datasets with
$\text{0.96}$
and 25 million samples. After obtaining the wrench matrix of the disc-magnet magnetic levitation actuator, the weighted pseudoinverse commutation law is adopted to decouple the system. Experimental validation shows multidisc-magnet mover control resolutions of
$\pm$
10
$\mu$
m and
$\pm$
10 mdegrees in the translational and rotational axes, respectively, with a processing time of 4.1
$\mu$
s. Full operating range sinusoidal responses demonstrate the capability of dynamic motion controls and motion ranges in all axes.