Identificação de sistemas aplicada a motores de indução trifásicos

Abstract

Three-phase induction motors are responsible for the highest energy consumption in the industry, serving as the main machines for converting electrical energy into mechanical energy. They are widely used due to their robustness and applicability across different operational scenarios. Understanding the physical behavior of these machines is essential for process control and monitoring, such as fault detection. For this, it is important to develop representative models that accurately capture the nonlinearities that typically appear only in real-world applications. In this context, this work presents a comparative analysis of system identification methods applied to the modeling of three-phase induction motors, using both classical approaches and machine learning-based techniques. Experimental data from the operation of threephase induction motors were collected on a test bench, considering torque variations from 0.5 to 4.0 N.m, and used to train and validate linear models (ARX and ARMAX) and nonlinear models (NARX) combined with different intelligent techniques, such as artificial neural networks, wavelet networks, and binary trees. The ARX and ARMAX models showed good performance in scenarios with low nonlinearity, with the fourthorder ARMAX model achieving 96.9% accuracy. Being simple and computationally inexpensive, these models are useful in well-defined industrial applications. Among the NARX models, the binary tree structure was the most accurate (97.1%) with the lowest standard deviation (0.112%), followed by the wavelet network (96.7%), both outperforming the linear models and other NARX configurations. The analysis highlights the importance of nonlinear models for accurately representing induction motors in complex scenarios, with the binary tree emerging as the best option. However, linear models remain viable when simplicity and interpretability are priorities. This work reinforces the value of system identification as a strategic tool for predictive maintenance, simulation, and digital twins, and suggests including fault data and mechanical variables in future studies to expand the practical applicability of the models.