CNC Precision Double: How High-Frequency Filtering Keeps Two Number Sets Honest

• 1
  Where the second number set comes from
• 2
  What high-frequency noise actually looks like at the axis
• 3
  What the filter does to the error term, and what it costs
• 4
  When filtering is the wrong answer
• 5
  How to verify a filter change did something real
• 6
  Matching the filter approach to the symptom

Every closed-loop axis runs two data paths. The controller issues a position command, usually as a 32-bit value with a fixed resolution. The encoder returns an actual position, and on modern drives that feedback is processed at higher internal resolution, often 64-bit floating point before it is scaled back to counts. So the same physical position exists twice in the system, at different numeric resolutions. The gap between them is what the drive tries to close.

High-frequency filter settings sit inside that loop. They attenuate encoder signal content above a set corner frequency, typically a few kilohertz on a mill axis. The intent is narrow: strip electrical noise that the drive would otherwise read as real motion. The filter does not improve mechanical accuracy. It prevents the loop from reacting to something that is not mechanically there.

This distinction matters when you talk about CNC precision double. People sometimes assume a higher-resolution number set automatically produces a better part. It does not. It produces a more sensitive loop. Sensitivity only pays off if the signal feeding it is clean and the mechanics can actually resolve the commanded increments.

• 1
  Command pathFixed-point resolution, set by the controller parameter.
• 2
  Feedback pathEncoder counts scaled up, filtered, then compared.
• 3
  Error termThe difference the drive amplifies into motor current.

Encoder noise is rarely a clean sine wave. On a real machine it shows up as short bursts that coincide with something else switching. A spindle drive accelerating, a coolant pump starting, a servo brake releasing, a welding cell on the same plant feed. The bursts last microseconds but repeat at kilohertz rates, which is why they land right in the band the filter is designed to cut.

The mechanical symptom is easy to misread. Operators describe it as a rough finish on one quadrant of a bore, or a faint ripple on a face milled at high feed. It looks like a spindle problem. It is more often a feedback problem, because the drive is injecting small current corrections that never correspond to a real tool movement.

Grounding and shielding cause more of these faults than filter settings do. A shield grounded at both ends creates a loop that picks up plant noise. A shield grounded at one end only, usually the drive end, behaves better. Cable routing matters too: encoder cable run parallel to a spindle power cable for 2 m will couple noise regardless of filter tuning.

• 1
  Switching eventsSpindle, pump and brake contactors in the same cabinet.
• 2
  Ground loopsShield bonded at both ends couples plant noise.
• 3
  Cable routingParallel runs with power cable over long distances.
• 4
  Worn encoderContaminated disc or failing bearing adds real signal errors.

A low-pass filter on the feedback path trades bandwidth for stability. Set the corner frequency high and the drive follows fast commands but also follows noise. Set it low and the axis runs quiet but lags on quick reversals, which shows up as rounded corners on a pocket and overshoot on a fast approach. Most builders land somewhere between 1 kHz and 4 kHz for milling axes and adjust from there.

There is a second-order effect worth knowing. Filtering the feedback delays the error signal, and a delayed error signal in a high-gain loop pushes the system toward oscillation. So adding filtering can, in some setups, make a marginal axis worse. That is why you tune gain and filter together, not one after the other.

Notch filters handle a different case. When one specific frequency is the problem, usually a spindle harmonic or a structural resonance, a narrow notch removes it with far less phase penalty than a broad low-pass. The catch is that the notch must track the disturbance. A fixed notch tuned to 420 Hz does nothing once the spindle speed changes and the harmonic moves.

Filtering cannot fix a mechanical fault. If a ball screw has 15 μm of backlash, or a thrust bearing is failing, the axis will show position error no matter how the drive is tuned. The filter only changes how the loop responds to the signal. It does not change what the mechanics can hold.

Filtering also cannot recover resolution that was never there. If the encoder provides 1 μm counts on a 40 mm pitch screw, the loop cannot resolve 0.005 mm regardless of how the numbers are processed downstream. High-resolution feedback is a hardware question. Software scaling just moves where the rounding happens.

The last boundary is thermal. A drive tuned tight with light filtering runs the motor harder, and on a long roughing cycle that shows up as axis drift. The mismatch between commanded and actual position grows as the screw warms. Parts that measure fine at 08:00 may drift by mid-afternoon. That is a thermal compensation problem, not a filter problem, and adding filtering only masks the drift while it hides the noise.

• 1
  BacklashMechanical lost motion; filtering cannot remove it.
• 2
  Encoder resolutionA hardware limit, not a tuning parameter.
• 3
  Thermal growthNeeds compensation, not attenuation.

The only honest test is a cut, measured. Pick a part feature that is sensitive to axis behavior, a bored bore, a circular interpolated pocket, a face milled at high feed. Cut it before and after the change. Measure roundness on a CMM or with a spindle-mounted indicator if the tolerance allows. A change in surface finish without a change in measured geometry is not proof of anything.

For the servo side, most drives log following error. Capture that trace during a representative move, ideally a rapid reversal followed by a slow feed. A clean loop shows a smooth error curve that returns to near zero. A noisy loop shows a jagged error with high-frequency content. Compare the traces, not just the peak value.

At GreatLight, this is part of how we qualify a process before a run. With 127 high-precision CNC machines across three plants, including 16 simultaneous 5-axis machining centers and 16 mill-turn centers, a tuning change made on one machine has to be proven on that machine. We hold ±0.005 mm and inspect 100% of parts before shipment, so a filter change that helps one axis and hurts another would show up in the inspection data.

• 1
  Cut and measureRoundness, flatness, or surface finish on a real feature.
• 2
  Log following errorCompare traces before and after, not peak values.
• 3
  Run a warm cycleRe-check after two hours of cutting to catch drift.