Learned crosstalk cancellation
Learning the leak: neural crosstalk cancellation for magnetic position sensing
When the coils that move a levitating puck leak into the sensors that track it, a network that also sees the coil currents cut the position error to 0.24 mm from 0.43 mm with the usual fix. The gap grows as the coils are driven harder.
Every result here comes from simulations of our own simplified models, run with fixed random seeds so they repeat exactly. Nothing has been measured on hardware yet. We report what came out, including where our approach did not win.
The question
Our levitating planar mover design (FT-100) senses its puck with Hall sensors placed under the very coils that push it. Every amp in a coil shows up in the sensor readings. The usual fix is to measure each coil's leakage once with no puck present and subtract it. But the leakage changes when the puck's magnet is nearby, and the sensors saturate. Can a small network do better?
What we did
We modelled 36 coils on a 20 mm grid with a sensor under each, a 61 mm magnet puck, leakage that grows by up to 25 % near the puck, soft sensor saturation, per-sensor gain and offset spread, and 0.15 mT of noise. We compared five ways to estimate position on 10,000 test positions, at drive levels from gentle to hard:
Symbols
- what Hall sensor s reads (mT)
- the puck magnet's field at the sensor, for puck position p
- coupling from coil c to sensor s, a distance d apart (mT per A)
- current in coil c
- how close the puck's magnet is to coil c
- extra leak near the puck (up to 25 %)
- the sensor's range, where it saturates softly
- this sensor's own gain and offset
In plain words
What sensor s reads: the puck's own field at that sensor, plus every coil's leak. A leak is the coupling K for the distance d to that coil, times its current I, made up to 25 % stronger when the puck's magnet is near the coil (n_c, from 0 to 1). The total is softly clipped by the sensor's range S. Each sensor also has its own gain g and offset o, and noise ε.
Symbols
- the measured leakage: 36 sensors × 36 coils, fitted by least squares with no puck present
- readings with the measured leak subtracted
- small networks (tanh hidden layers) that turn readings into a position
- the estimated puck position (x, y)
In plain words
The usual fix (B) measures each coil's leak once with no puck present, subtracts it, and hands the cleaned readings to the position network. The hybrid (E) subtracts the same measured leak, then gives a network both the cleaned readings and the currents, so it can learn what subtraction misses: the extra leak near the puck, saturation and sensor spread.
| Method | |
|---|---|
| A | Position network, calibrated with coils off |
| B | A plus measured leakage subtracted (usual fix) |
| D | Network trained with coils on, no current inputs |
| C | Network with readings and coil currents |
| E | Leakage subtracted, then a network with the currents (hybrid) |
What we found
Ignoring the coils is not an option: a network calibrated with the coils off was 2.76 mm out once they were driven. A network trained with the coils on but without knowing the currents got 1.35 mm. Subtracting the measured leakage brought it to 0.43 mm. The network that also sees the currents reached 0.26 mm, and the hybrid (subtract first, then learn what is left) reached 0.24 mm. With the coils off entirely, the same sensing gets 0.21 mm. That is the floor.
The advantage depends on how hard the coils work. In the gentlest band (0 to 0.25 A) subtraction gave 0.28 mm and the hybrid 0.20 mm; in the hardest (0.55 to 0.8 A), 0.64 mm against 0.33 mm. Subtraction assumes the leak is the same everywhere. It isn't, and the error from that assumption grows with the current.
What it means
Physics first, learning second works well: subtracting the part of the leak we can measure leaves the network a smaller, smoother problem. The network matters most exactly when the stage works hardest, with heavier loads and faster moves, which is when position accuracy matters most.
The same network is small enough for the stage's own microcontroller (study 3), and it is the network we made event-driven in study N2.
Limits: a simplified field model, not a field solve of the real design; one simulated unit; the leakage effect near the puck is assumed (+25 %), and a real stage could show more or less.