Synthetic Data-Driven Detection of Broken Rotor Bars in Induction Machines Under Adjusted Noise Level

This paper proposes a synthetic data-based approach for diagnosing broken rotor bars in induction machines, aiming to bridge the gap between synthetic and experimental datasets in broken rotor bars analysis. By comparing the current differences between simulated and experimental data, harmonized noise is generated and integrated into simulated current signals to further train tree-based machine learning algorithms for rotor health classification. High-frequency noise components and harmonics are isolated through Fourier series decomposition to mathematically model the noise. In a first study, the mathematically modeled noise is substantially added to simulated signals, enhancing diagnostic accuracy by more than 20% compared to algorithms trained solely on simulated signals. In a second study, different levels of noise are progressively added to simulated signals, showing that higher noise percentages enhanced model generalization, with accuracy reaching its peak at 100% noise inclusion. This work emphasizes the importance of noise in effectively detecting broken rotor bars in real-world scenarios, highlighting that industrial simulations must account for noise to be accurate if they rely solely on synthetic data.