A virtual sensor for backlash in robotic manipulators

Eliana Giovannitti, Sayyid Shahab Nabavi, Giovanni Squillero, Alberto TondaJ. Intell. Manuf 2022 · DOI

Abstract

This paper presents a non-invasive "virtual sensor" methodology for estimating mechanical backlash in industrial robotic joints. By analyzing motor encoder signals during specific excitation cycles and employing the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), the system identifies vibration signatures indicative of gear wear. Experimental results across simulation and real-world industrial environments demonstrate high accuracy and the ability to distinguish between healthy and degraded components without additional hardware, enabling cost-effective Condition-Based Maintenance (CBM).

Introduction

In industrial robotics, precision is paramount. However, mechanical components are subject to wear, leading to backlash—the clearance or "play" between mating gear teeth. While a minimal amount of backlash is necessary for smooth operation and lubrication, excessive wear widens this gap, resulting in positioning inaccuracies, unwanted vibrations, and potential mechanical failure.

System architecture showing the motor, encoder, and the joint where backlash occurs.

Traditionally, measuring this degradation requires manual inspection with dial indicators, which necessitates significant production downtime, or the installation of expensive external sensors like accelerometers. This research proposes a "software-only" solution that leverages existing high-resolution motor encoders to detect the invisible signature of backlash.

System Architecture and the Backlash Problem

The primary challenge in detecting backlash from motor data is that the encoder is typically located on the motor shaft, separated from the output link by the joint's transmission system. As shown in the schematic below, the backlash occurs within the joint itself, making it difficult to observe directly from the motor's perspective.

System architecture showing the motor, encoder, and the joint where backlash occurs.

When the robot changes direction or moves through a specific cycle, the transition across this mechanical gap causes tiny, damped oscillations in the motor speed. These oscillations serve as a physical "signature" of the backlash level.

Methodology

The proposed virtual sensor operates through a multi-stage pipeline designed to extract these subtle signals from noisy industrial data.

1. Excitation Cycle

The robot performs a controlled test movement, typically on a horizontal axis. By utilizing gravity, the researchers ensure the gears "fall" through the backlash gap at a predictable moment, creating a repeatable impact signature that is easier to isolate from normal operational noise.

2. Signature Modeling and Optimization

The system uses a mathematical model, $h(t)$, to represent the expected sequence of damped oscillations. The amplitude ($A$) of these oscillations is the critical parameter, as it correlates directly with the physical width of the backlash gap.

Because the motor signal is inherently noisy and the optimization landscape is non-convex (making it difficult for standard algorithms to find the global optimum), the authors employed the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This evolutionary algorithm is robust enough to fit the mathematical model to the residual speed signal even under low signal-to-noise conditions.

3. Virtual Mapping

Once the vibration amplitude ($A$) is identified, a cubic polynomial function maps this value to the actual physical backlash ($\delta$), providing a tangible metric for maintenance teams.

Experimental Validation

The methodology was validated using four distinct data sources to ensure robustness:

  • High-Fidelity Simulation: A Matlab/Simulink model of a robotic joint was used to test 10 different backlash levels.
  • Noisy Data Testing: Uniform white noise was added to simulations to mimic electronic interference found in factory environments.
  • Long-term Test Bench: A COMAU robot joint was operated under high stress for eight months, with manual ground-truth measurements taken every three weeks as wear naturally progressed.
  • Industrial Production Robots: The algorithm was deployed on two active robots in a Brazilian manufacturing plant—one healthy and one with known mechanical issues.

Key Results and Discussion

The virtual sensor demonstrated remarkable precision. In noisy simulations, the mean relative error for vibration amplitude estimation was typically below 10%, and frequently as low as 1-2%. The maximum absolute error in physical backlash estimation was approximately 0.0004 rad, a resolution sufficient for industrial maintenance standards.

In the industrial field test, the system successfully distinguished between the two robots. "Robot A" (healthy) returned an amplitude of ~0.02, while "Robot B" (worn) returned ~0.22. This clear differentiation proves the system's efficacy in uncontrolled, real-world environments.

Conclusion

This research provides a scalable, cost-effective tool for robotic maintenance. By transforming existing data into a diagnostic "virtual sensor," manufacturers can move from reactive repairs to proactive, condition-based maintenance. Future work will focus on refining the calibration process for different robot models and exploring passive monitoring techniques that do not require dedicated test cycles.

Summary updated

BibTeX
@article{giovannitti2022virtual,
  author = {Eliana Giovannitti and Sayyid Shahab Nabavi and Giovanni Squillero and Alberto Tonda},
  title = {{A virtual sensor for backlash in robotic manipulators}},
  journal = {J. Intell. Manuf},
  year = {2022},
  doi = {10.1007/s10845-022-01934-z},
}
Figure