Data-Driven Intermittent Fault Diagnosis: From Data Acquisition to CNN-Based Classification in Synchronous Generators

Condition monitoring is essential for maintaining the reliability of electrical machinery, particularly synchronous generators used in power production, maritime applications, military operations, and laboratory testing. Stator inter-turn defects are a significant issue due to their potential to cause efficiency decline and machine failure. Although existing test beds primarily focus on permanent faults, they lack the capability to replicate intermittent faults. This study presents a Fault Injection Unit (FIU) programmed with LabVIEW to generate intermittent faults with durations of $0.5 \mathrm{~s}, 0.2 \mathrm{~s}$, and 0.1 s, alongside permanent faults, under various load conditions, 0.5 A and 1 A. A 1D Convolutional Neural Network (CNN) trained on permanent fault data is evaluated for intermittent fault classification. The results show that while the model excels at identifying permanent faults, its accuracy decreases for intermittent faults, particularly with shorter fault durations. FFT analysis reveals that longer intermittent faults increase frequency ripple content at the third harmonic (150 Hz). The impact of preprocessing methods, including Short-Time Fourier Transform (STFT) and wavelet analysis, on classification accuracy is also examined, highlighting the need for specialized processing techniques to improve intermittent fault detection.