Partial-discharge (PD) monitoring is essential for assessing the insulation condition of generator stator bars. Conventional methods are susceptible to electromagnetic interference and are difficult to deploy in confined stator geometries. Fiber-optic acoustic detection technology offers strong immunity to electromagnetic interference and is suitable for the narrow and high-interference environment of stator bars, but it cannot directly provide discharge magnitude information. Therefore, in this study, fiber-optic acoustic detection technology was employed to acquire partial discharge acoustic signals from stator bars, and a mandrel-type fiber-optic acoustic sensor was developed, with PD tests performed on full-scale stator bars with internal defects. Meanwhile, considering the complex temporal characteristics of PD acoustic signals, a hybrid neural network—Transformer–convolutional neural network–long short-term memory (Transformer–CNN–LSTM)—was constructed for long-term time-series modeling to establish the mapping between acoustic signals and discharge magnitude intervals. The results indicate that fiber-optic acoustic detection enables sensitive and stable detection of weak PD acoustic signals. Phase-resolved PD (PRPD) patterns from the proposed system align with the discharge characteristics of internal defects, with the acoustic signal showing a phase lag relative to the electrical PD signal. The hybrid model achieved an overall interval estimation accuracy of 96.6%, outperforming CNN and CNN-LSTM models, with accuracies of 100% and 99.4% for discharge magnitude intervals below 100 pC and above 2000 pC, respectively.
Hu et al. (Wed,) studied this question.