⚡ electrical engineering

Performance-based Adaptation Termination for Preventing Parameter Drift in Adaptive Vibration Suppression

This paper proposes a computationally efficient, RMS-based stopping criterion for adaptive vibration control that prevents parameter drift and preserves suppression performance by freezing adaptation once satisfactory attenuation is sustained, as validated through simulations and experiments on a cantilever-beam platform.

Original authors: Juan Augusto Paredes Salazar, Ankit Goel

Published 2026-08-04

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Juan Augusto Paredes Salazar, Ankit Goel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to balance a broom on its hand. At first, the broom is wobbling wildly, so the robot's brain frantically adjusts its grip, learning from every tiny shake to find the perfect balance. This is the world of adaptive control: a type of smart engineering where machines learn on the fly to cancel out unwanted vibrations, like the shaking of a bridge in the wind or the wobble of a satellite's solar panels. The goal is to use data from sensors to automatically tune the machine's behavior until the shaking stops.

However, there is a tricky problem with this "learning forever" approach. Once the robot has successfully balanced the broom and the shaking has mostly stopped, the robot might get confused. Because the broom is now so steady, there is very little new information for the robot to learn from. If the robot keeps trying to "learn" in this quiet state, it might start guessing wildly, tweaking its settings based on tiny sensor glitches or background noise. This is called parameter drift. It's like a student who, after acing a test, keeps changing their answers just because they are bored, eventually ruining a perfect score. In engineering, this can cause a machine that was working perfectly to suddenly become unstable and start shaking again.

This paper tackles that exact problem. The researchers, Juan Augusto Paredes Salazar and Ankit Goel, propose a clever "stop learning" switch for these adaptive machines. Instead of letting the robot guess forever, they suggest a rule: once the vibration is low enough for a specific amount of time, the robot should freeze its brain and stop updating its settings. They tested this idea on a flexible metal beam that was being shaken by a motor. By using a simple mathematical tool called a "Root-Mean-Square" (RMS) calculator—which acts like a sensor that measures the average energy of the shaking—they found a way to tell the difference between "good, quiet stability" and "dangerous, noisy guessing."

The team showed through both computer simulations and real-world experiments that without this stop switch, the controller kept changing its mind even after the vibrations were gone, eventually making the beam wobble again. But when they turned on their new RMS-based stop switch, the controller learned quickly to stop the shaking, and then immediately froze its settings. This kept the beam steady and prevented the "drift" that usually ruins the performance. The result is a smarter, more reliable way to control vibrations that doesn't require complex new math, just a simple rule to know when to stop learning and start resting.

The Story of the Shaking Beam

To understand why this matters, picture a long, thin aluminum beam sticking out from a wall, like a diving board. In the real world, this beam is a "flexible structure," meaning it's not rigid; it wants to wiggle. In the lab, the researchers clamped one end of this beam (0.5 meters long) and attached two small motors to it. One motor was the "bad guy," shaking the beam to create a disturbance, while the other motor was the "good guy," trying to push back and stop the shaking.

The challenge was that the "good guy" motor wasn't right next to the sensor measuring the shake. This is called a noncollocated setup, which is like trying to steer a car by looking in the rearview mirror instead of through the windshield. It makes the control problem much harder because the feedback is delayed and tricky. The researchers used a smart algorithm called Retrospective Cost Adaptive Control (RCAC). Think of RCAC as a detective that looks back at the last few seconds of data to figure out how to adjust the motor to cancel out the next shake.

The Problem: When Learning Becomes a Bad Habit

In their experiments, they set the "bad guy" motor to shake the beam at a frequency of 20 Hz (20 times a second). When the RCAC controller was turned on, it did a great job. It quickly learned the pattern of the shake and pushed back, reducing the vibration at the tip of the beam significantly.

But here is where the trouble started. Once the beam was calm, the "bad guy" motor was still running, but the beam wasn't moving much. The controller, still in "learning mode," kept trying to tweak its settings. Because the beam was so quiet, the tiny bit of noise from the sensors (like a faint hum or a digital glitch) looked like a big signal to the confused controller. The controller started making tiny, unnecessary changes to its internal numbers (its parameters).

Over time, these tiny changes added up. The controller drifted away from the perfect settings it had found. It was like a musician who, after hitting the perfect note, starts fiddling with the tuning pegs unnecessarily and ends up playing out of tune. In the simulations, this drift caused the vibration suppression to get worse, and in the worst cases, it threatened to make the system unstable. The paper explicitly argues against the idea that the controller should just keep learning forever; they show that continuing to adapt after the job is done is actually harmful.

The Solution: The "RMS" Stop Sign

To fix this, the authors introduced a simple but powerful idea: Performance-Based Adaptation Termination.

Instead of looking at the vibration at a single instant (which might be zero just because the wave happened to cross the middle line), they looked at the Root-Mean-Square (RMS) of the vibration. If you imagine the vibration as a wave, the RMS is like measuring the average height of the entire wave over a short period. It tells you how much "energy" the shaking has, rather than just its momentary position.

They set up a rule with two parts:

The Threshold: They decided that if the RMS energy of the vibration drops below a certain level (0.05 in their specific units), the beam is "quiet enough."
The Memory: They didn't just check this once. They used a "forgetting factor" (a number called $\gamma$ ) to calculate a moving average. This meant the system had to stay quiet for a sustained period (an effective window of 500 samples in the simulation, or 10 samples in the experiment) before it decided to stop.

Once the RMS value stayed below that threshold for long enough, the system hit the "pause" button. It froze the controller's internal settings. The controller stopped trying to learn and simply held its ground with the settings it had just found.

What They Found

The results were clear and convincing.

In the computer simulations:
The researchers modeled the beam using a "lumped-parameter" model (a simplified version of the physics). They ran the beam with a 20 Hz disturbance.

Without the stop switch: The controller kept updating its settings. The graph of the controller's parameters (the numbers inside the brain) started to wander off. Eventually, the vibration suppression got worse, and the beam started shaking again.
With the stop switch: The controller learned fast, the vibration dropped, and the RMS value fell below the line. The system froze the settings. The beam stayed calm, and the controller's numbers stayed exactly where they were.

In the real-world experiment:
They built the actual aluminum beam setup described earlier, using a laser vibrometer to measure the tip's movement with extreme precision. They used a dSPACE MicroLabBox II computer to run the control loop in real-time.

Without the stop switch: Just like in the simulation, the controller kept tweaking itself after the shaking stopped. The vibration suppression degraded, and the system became less effective.
With the stop switch: The controller adapted, the vibration dropped, and the RMS value hit the target. The system froze the parameters. The beam remained stable, and the vibration stayed suppressed.

The paper is very clear about the limits of these findings. They proved through simulation and measured through experiment that this method works for this specific type of single-mode vibration suppression. They did not claim it solves every possible problem in the universe, nor did they claim it works for multi-mode structures (where many different frequencies shake at once) without further testing. They explicitly state that future work is needed to see if this holds up under broader conditions.

Why This Matters

The beauty of this solution is its simplicity. It doesn't require the engineers to build a complex new model of the beam or to add heavy, slow math to the computer. It just adds a tiny bit of extra calculation (recursively updating the RMS) that is almost free in terms of computing power.

It's like giving the robot a "good job" certificate. Once the robot has done the job well for long enough, the certificate tells it, "Stop trying to be perfect; you're already perfect. Just hold the pose." This prevents the robot from overthinking and messing up a good thing.

The authors conclude that this method is a practical, low-cost way to prevent the "parameter drift" that plagues adaptive systems. By using a simple performance monitor, they can keep the benefits of learning (fast adaptation when things are chaotic) while avoiding the downsides (drifting when things are calm). It's a small change in the rules that keeps the machine stable, safe, and effective.

Technical Summary: Performance-Based Adaptation Termination for Preventing Parameter Drift in Adaptive Vibration Suppression

Problem Statement
Adaptive vibration control systems for flexible structures face a persistent practical limitation known as parameter drift. While adaptive controllers effectively attenuate disturbances by tuning parameters based on measured data, a critical issue arises once satisfactory vibration suppression is achieved. At this stage, the oscillatory response diminishes, reducing the "persistence of excitation" required for stable parameter estimation. In the presence of measurement noise or unmodeled dynamics, the adaptation problem becomes ill-conditioned. Consequently, the controller may continue to update parameters unnecessarily, leading to gradual drift, performance degradation, or even closed-loop instability. Traditional mitigation strategies, such as projection operators, normalization signals, or leakage terms ( $e$ - and $\sigma$ -modifications), often alter the nominal adaptive law, introduce additional tuning parameters, or create trade-offs between robustness and disturbance rejection performance.

Methodology
This paper proposes a performance-based adaptation termination mechanism that mitigates parameter drift without modifying the underlying adaptive update law. The core methodology integrates a computationally efficient stopping criterion with Retrospective Cost Adaptive Control (RCAC).

RMS-Based Stopping Criterion: Instead of relying on instantaneous error values (which are unsuitable for oscillatory signals), the method utilizes a recursively computed Root-Mean-Square (RMS) of the performance variable ( $z_k$ ).
- An Exponentially Weighted Moving Average (EWMA) is applied to the squared performance variable to calculate $z_{rms, k}$ .
- The forgetting factor $\gamma$ is related to an effective moving-window length ( $n_{samp}$ ), allowing the metric to smooth transient noise while tracking sustained vibration energy.
- A user-defined threshold ( $z_{rms, stop}$ ) is established.
Adaptation Termination Logic:
- As long as $z_{rms, k} \geq z_{rms, stop}$ , the RCAC algorithm continues to update controller parameters ( $\theta_k$ ) and the covariance matrix ( $P_k$ ) using standard recursive least-squares (RLS) equations.
- Once $z_{rms, k}$ falls below the threshold and remains there, parameter updates are frozen ( $P_k = P_{k-1}$ , $\theta_k = \theta_{k-1}$ ).
- The controller effectively transitions from an adaptive mode to a fixed-gain mode, preserving the parameters that achieved the desired attenuation.
Implementation: The approach is implemented on a cantilever beam platform with a noncollocated actuator configuration (actuator and sensor at different locations). Validation is performed through both numerical simulations using a lumped-parameter model (LPM) and real-time closed-loop experiments.

Key Contributions
The paper identifies three primary contributions:

Performance-Based Termination: A mechanism to halt adaptation once sustained vibration attenuation is achieved, preventing parameter drift without altering the nominal adaptive update law (unlike projection or leakage methods).
Recursive RMS Metric: The development of a computationally lightweight, recursively calculated RMS metric based on EWMA. This provides an intuitive tuning guideline via the effective window length ( $n_{samp}$ ) and is specifically suited for oscillatory vibration responses that do not converge to zero.
Experimental and Simulation Validation: Demonstration of the method's efficacy on a flexible structure with noncollocated sensing and actuation, showing that the approach preserves transient benefits while ensuring bounded parameter estimates.

Results

Numerical Simulations: In simulations without the stopping criterion, the controller continued to adapt after vibration suppression, leading to gradual parameter drift and eventual degradation of vibration suppression performance. With the RMS-based criterion enabled, adaptation ceased once the RMS threshold was met, keeping parameters bounded and maintaining the achieved attenuation.
Experimental Validation: Closed-loop experiments on the aluminum cantilever beam replicated the simulation findings. Without the criterion, continued adaptation led to performance degradation. With the criterion, the system maintained stable closed-loop performance, and the controller parameters remained bounded. The results confirmed that the method prevents destabilizing parameter evolution while retaining the adaptation benefits necessary for initial vibration suppression.

Significance and Claims
The authors claim that this approach offers a computationally efficient and minimal-tuning solution to parameter drift. Key advantages highlighted include:

No Model Requirement: The method does not require an explicit plant model.
Low Computational Overhead: The recursive RMS calculation requires negligible additional effort, making it suitable for high-rate embedded control systems.
Preservation of Performance: Unlike methods that introduce steady-state bias or slow convergence (e.g., leakage terms), this approach retains the transient benefits of adaptation and the final disturbance rejection performance.
Generality: While integrated with RCAC in this work, the mechanism is agnostic to the specific adaptive law and can be applied to a broad class of gradient-based or model-reference adaptive controllers.

The paper concludes that a simple performance-based monitoring mechanism can effectively prevent parameter drift in adaptive vibration control, ensuring that the system remains stable and effective once the primary control objective is met. Future work is suggested to investigate analytical guarantees for boundedness under broader excitation conditions and to extend the approach to multi-mode flexible structures.