How Robust Are Neural Networks Modeling Electric Motors?

Created in December 02, 2022

2022   ·   robotics   deep-learning   time-series   robustness   ·   research

Why Motors, and Why Robustness

Every robot ultimately moves the world through electric motors. If you want learning-based low-level control — torque prediction, temperature estimation, fault detection — you increasingly reach for neural networks that model motor dynamics from time-series signals. But a model that is accurate on clean benchmark data is not the same as a model you can trust inside an actuator loop. Sensor noise, distribution shift, and hardware faults are the norm, not the exception.

This work — “Robustness of Neural Networks used in Electrical Motor Time-Series,” presented at the NeurIPS 2022 RobustSeq Workshop — asks a pointed question: how do these models behave once you stop feeding them pristine data?

The Setup

We studied neural network models across three complementary electric-motor datasets:

  • Motor dynamics / denoising / speed-torque — the core task of characterizing motor behavior.
  • Temperature — thermal dynamics, which govern safe operating limits.
  • Broken bars — a classic fault-detection setting for degraded machines.

Across these, we probed how models trained on nominal data respond to adversarial perturbations, noise, and real-world variation — the conditions an actuator actually sees in the field.

What It Means for Robotics Actuators

The reason this sits in my robotics work and not just an ML-theory corner: a motor controller is a safety-critical component. If a learned dynamics or fault-detection model degrades ungracefully under noise or a partial hardware fault, the failure propagates straight into the robot’s motion. Understanding where these models are brittle — and building in robustness — is a prerequisite for putting learning-based control anywhere near real actuators.

This line of work traces back to my Ph.D. at CentraleSupélec, where I modeled the dynamics of heavy electrical motors with neural networks (funded by Schneider Electric), and it feeds directly into how I think about learning-based low-level control today.

Code, datasets, and pretrained weights: github.com/sagarverma/robust-motor.