Um estudo comparativo de robustez de modelos de aprendizado profundo para o diagnóstico de barras quebradas no rotor utilizando séries temporais de corrente do estator com ruído e dados incompletos

Abstract

Three-phase induction motors are widely deployed in industrial environments and are subject to mechanical and electrical degradation mechanisms. Among the relevant failure modes, broken rotor bars are recurrent and may lead to torque reduction, increased vibration, and thermal stress, potentially evolving into more severe damage when the fault is not diagnosed at an early stage. In this context, Motor Current Signature Analysis (MCSA) stands out as a non-invasive and cost-effective diagnostic strategy, since it relies exclusively on stator current measurements, avoids additional sensor instrumentation, and enables continuous online monitoring. This work presents a comparative investigation of the effectiveness and robustness of three deep learning architectures for automatic rotor-condition classification using only stator current time series: Long Short-Term Memory (LSTM), InceptionTime (IT), and Temporal Signal Transformer (TST). The models were trained and validated on a publicly available experimental dataset from IEEE DataPort comprising five rotor integrity levels (0 to 4 broken bars) under multiple load conditions. To ensure a fair comparison, a unified protocol was adopted for preprocessing, windowbased segmentation, training configuration, and evaluation metrics. The results indicate that LSTM achieved the best overall performance, reaching 98% accuracy, a weighted F1-score of 0.978, and Cohen’s Kappa of 0.973, outperforming IT and TST in the nominal setting. In addition, robustness analyses under controlled degradations in the test set further highlighted the stability of LSTM: in the presence of Additive White Gaussian Noise (AWGN), accuracy remained above 90% even under the extreme 0 dB SNR condition, and under incomplete data, performance remained around 88% with 30% missing samples. IT exhibited competitive results with gradual degradation under moderate perturbations, whereas TST, although effective in clean and moderately degraded scenarios, showed greater sensitivity under extreme noise and missing-data conditions. These findings provide evidence that deep learning-assisted MCSA is a feasible and reliable solution for multiclass diagnosis of broken rotor bars (BRB) severity using stator current alone, with direct potential for low-cost predictive maintenance. Among the evaluated approaches, recurrent architectures emerged as particularly robust and promising for deployment in industrial environments subject to noise and acquisition faults.