Stick-slip step prediction
Where the gain really comes from in stick-slip positioning
Predicting a piezo stage's step size from the drive alone was 37 % off. Feeding back the recently measured step brought it to 5.4 %, and a network added only a little more (5.2 %). The feedback, not the network, saved the time.
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
A piezo stick-slip positioner moves in steps of about 2 µm whose size depends on friction. Friction changes along the travel, with humidity and with wear. To make a move, the controller guesses how many steps it needs, runs them, reads its position sensor and corrects. A better guess means fewer correction passes. We expected a network that knows the position along the travel to guess best.
What we did
We built a physical model of the stick-slip cycle (slow ramp, 45 µs snap-back, static and kinetic friction, 35 N preload, 20 g slider), with friction that ripples along the 10 mm travel and a hidden humidity and wear factor of ±15 %. Simulated steps ranged from 0.16 to 2.13 µm. We compared three predictors on 5,000 test cases, then ran 300 closed-loop 1 mm moves with each.
Symbols
- the piezo's stroke for drive voltage V and step rate f; the slider rides it during the slow ramp
- how long the piezo takes to snap back
- slider and piezo velocities
- slider mass and preload
- kinetic and static friction
- this unit's friction offset, and the hidden humidity and wear factor (0.85 to 1.15)
- the net step: stroke minus back-slip
In plain words
During the slow ramp, friction carries the slider forward by the piezo's full stroke a. During the 45 µs snap-back, the piezo accelerates faster than friction can hold the slider, so the slider slips and is dragged back while the piezo moves, then slides to a stop. The net step is the stroke minus that back-slip. More friction (μ) means more drag and a smaller step. μ depends on the unit (u), on humidity and wear (h, hidden) and on the position along the travel (x).
Symbols
- predicted size of the next step (µm)
- average step the position sensor saw on the last move; it carries the hidden friction
- position along the 10 mm travel
- fitted coefficients and network weights
In plain words
Three ways to guess the next step size ŝ. P1 assumes it is proportional to voltage. P2 fits a curve to voltage and step rate, and also uses s_recent, the average step the position sensor saw on the last move, which carries the hidden friction. P3 is a small network given the same inputs plus the position x along the travel.
What we found
A step proportional to voltage, as a datasheet would suggest, was 37.5 % off: the backward slip during the snap is nearly fixed, so small drive voltages make disproportionately small steps. A linear calibration that includes the step size measured on the last move got 5.4 %. The network got 5.2 %.
| Predictor | Correction passes (mean) | Passes (95th pct) | Time (ms) |
|---|---|---|---|
| Step proportional to voltage (datasheet style) | 3.56 | 5 | 722 |
| Linear calibration with the recent measured step | 2.29 | 3 | 560 |
| Small network with position along the travel (ours) | 2.25 | 3 | 561 |
In closed loop, feeding back the measured step cut the mean correction passes from 3.56 to 2.29 and the move time from 722 to 560 ms. The network changed neither by much.
The chart shows why. The network smoothed over the fine friction ripple along the travel: from the raw position alone, with this much data, it did not learn a pattern that repeats every 0.7 mm.
What it means
Not every problem needs a network. Here, measuring and feeding back did almost all the work. That is useful to know before putting a network in a product: the cheaper fix was nearly as good.
The next step is a per-unit friction map built up as the stage runs. That is a lookup table rather than a network, and a natural job for on-chip learning on neuromorphic hardware.
Limits: a one-dimensional friction model with an assumed ripple; hidden factors drawn at random; closed loop without the fine mode a real stage would use for the last micrometre.