Comparative Analysis of LSTM and GBR for BLDC Motor Speed Estimation Under Noisy Conditions

Abstract

Estimation of BLDC motor speed plays a critical role in enhancing the efficiency and reliability of electric motors, particularly with the complexities of real-world operating conditions.In this work, we applied the techniques Long Short-Term Memory (LSTM) networks and Gradient Boosting Regressors (GBR) for estimating motor speed in environments characterized by noise and unpredictable changes.The dataset was generated from a numerical simulation of the BLDC motor dynamic model (Linix 45ZWN24-40), implemented in MATLAB/Simulink. This simulation-based approach was adopted to allow controlled introduction of Gaussian noise and abrupt torque transitions, enabling systematic evaluation of model robustness before future physical implementation. The input features consist of voltage, load torque, and motor parameters (B, Ls) varied across scenarios to represent real-world parameter uncertainty. The simulations conducted with the LSTM model resulted in a mean squared error (MSE) of 2580.12 and an R-squared value of 0.95. In contrast, the Gradient Boosting Regressor (GBR) achieved an MSE of 3150.87 and an R-squared value of 0.93. While GBR requires less time for training, LSTM consistently provided higher accuracy, particularly during the rapid variations in torque.This systematic comparison of two machine learning models offers practical insights for engineers tasked with developing motor control systems in unpredictable and dynamic environments.