In-Context Learning for Zero-Shot Speed Estimation of BLDC motors

TL;DR

The paper tackles sensorless speed estimation for BLDC motors, where model-based estimators struggle with nonlinearities and parameter uncertainty. It introduces an in-context learning framework using a transformer-based contextual filter trained offline on simulated BLDC trajectories to perform zero-shot speed estimation from electrical measurements. To bridge the sim-to-real gap and mitigate aliasing, the method conditions on measurement history and includes the previous speed estimate as input. Experimental validation on a real motor shows the transformer estimator outperforming a set of EKFs, particularly during startup and at low speeds, demonstrating practical viability and rapid deployment without per-motor retraining.

Abstract

Accurate speed estimation in sensorless brushless DC motors is essential for high-performance control and monitoring, yet conventional model-based approaches struggle with system nonlinearities and parameter uncertainties. In this work, we propose an in-context learning framework leveraging transformer-based models to perform zero-shot speed estimation using only electrical measurements. By training the filter offline on simulated motor trajectories, we enable real-time inference on unseen real motors without retraining, eliminating the need for explicit system identification while retaining adaptability to varying operating conditions. Experimental results demonstrate that our method outperforms traditional Kalman filter-based estimators, especially in low-speed regimes that are crucial during motor startup.

Figures (6)

• Figure 1: Experimental setup for a specific configuration. Note that, the motor and the inertial load are connected through a spindle.
• Figure 2: Speed response of different BLDC motor configurations when subjected to the same quadrature current reference profile.
• Figure 3: Examples of speed trajectories corresponding to disk configuration $S^{(2)}$ (first column) to $S^{(6)}$ (last column). True speed (dashed-black line) is compared against the speed estimate from the contextual filter (orange) and EKFs (blue) estimations. The last row displays the estimation error and standard deviation (shaded areas) for the specific configuration averaged over the $M=15$ experiments.
• Figure 4: Average estimation along the fixed reference experiments.
• Figure 5: Boxplot of the estimation error $\varepsilon$ for the specific testing configurations averaged over time and the experiments.

Theorems & Definitions (5)

• Remark 1: Maximum allowed sequence length
• Remark 2: On the need of $\omega_{e,0} \leq \omega_{e,\max}$
• Remark 3: Training the contextual filter
• Remark 4: Case with $T_s \ll T^{\mathrm{eval}}$
• Remark 5: Sensored closed-loop experiments