Decoupling Friction Compensation from Load Disturbance: The Truth About Servo Tuning through Spectral Analysis

During on-site commissioning in factory automation, we often encounter a frustrating scenario: when the motor runs at low speeds, the tracking error just won't stay within the target range. Many engineers' first instinct is to tweak the PID parameters or crank up the gain for friction compensation. But here's the catch—is that really helping? Sometimes, what we perceive as "stick-slip" caused by friction is actually tangled up with external load dynamics. If we don't clearly distinguish between the two sources, "blind" compensation often introduces unnecessary oscillation into the system.

Back to Basics: Why Do We Need to Separate the Error?

In control theory, nonlinear friction (like stiction or Coulomb friction) and load disturbance (such as variations in external load inertia or cutting forces) look quite similar as errors in the time domain. However, if we peel them apart and look at the physics, their characteristics are fundamentally different:

• Nonlinear Friction: Heavily dependent on position and velocity, typically manifesting as distinct jumps or hysteresis at velocity reversal points.
• Load Disturbance: Usually related to the system's dynamic state; it can be stochastic or superimposed on the periodicity of the load.

When tuning, the tracking error we see is just a superposition of both. To decouple them, we need to move into the frequency domain, where these phenomena leave very different "footprints" in the spectrum.

Non-Invasive Quantification: Power Spectral Density (PSD) of Current Commands

Since we can't always strip the machine down, we use internal "data monitoring" from the drive to find the answers. My go-to tool is analyzing the Power Spectral Density (PSD) of the current (or voltage) commands.

How to Interpret Spectral Analysis

When the system is in motion, we record a segment of current command data at a steady velocity and perform a Fast Fourier Transform (FFT). You'll find that nonlinear effects caused by friction usually show up as harmonics in the lower frequency range—specifically the fundamental frequency related to the motion and its multiples.

Key Insight: If the friction compensation model is designed correctly, those odd-order harmonic peaks in the PSD plot (corresponding to velocity polarity switching) will significantly decrease near the reversal points. Conversely, if load disturbance is the culprit, you'll see noise components across a wider bandwidth or at specific frequency points corresponding to the load's periodicity.

The Key Technique for Decoupling: Frequency Windowing in Disturbance Observers (DOB)

By designing a Disturbance Observer (DOB), we can categorize observed disturbances into a "low-frequency region" (where friction dominates) and a "high-frequency region" (where load uncertainty dominates). In practice, if you observe significant jitter in the output current command at higher frequencies, it usually suggests an issue with load parameter estimation, not friction compensation.

Practical Tuning Considerations

Warning: Never crank up friction compensation gains just to make the tracking error "look pretty" on a scope. This leads to high-frequency over-correction, known in servo control as "Hunting," which will prematurely wear out your mechanical transmission components, like ball screws.

The essence of tuning lies in distinguishing between "predictable friction" and "random load disturbance." By using current PSD as a non-invasive quantitative tool, you can clearly see whether your friction model parameters are actually reducing the nonlinear contribution. Once the nonlinear part is minimized, whatever remains falls under load disturbance—that’s when you should start looking into your Model Predictive Control (MPC) state observer or adjusting your filters to handle the residuals.

Automation control is both an art and a rigorous engineering science. If you break it down, the problems are always hidden in the fundamental frequency response. I hope the next time you're on the factory floor, you'll listen closely to what the system is telling you through its spectral analysis.