Detection of Partial Discharge Acoustic Signals in Converter Valve Halls Using Multivariate Time–Frequency Slices

Partial discharge serves as a critical early warning signal for the aging and potential failure of power equipment within converter valve halls. Acoustic detection has emerged as a promising approach for capturing this signal. However, identifying faint discharge sounds amid the complex interference environment of valve halls remains a significant challenge. To address this issue, this study proposes an innovative detection framework leveraging multiple time-frequency slices derived from both field-collected and laboratory-simulated discharge sound samples. First, a novel wavelet time-reassigned synchrosqueezed transform (WTSST) is introduced, which enhances time-frequency resolution through temporal dimension compression. This method effectively extracts electroacoustic pulses and generates time-frequency slices within frequency bands free from converter valve operational noise. Second, for converter stations with limited computational resources, a lightweight verification algorithm based on variational mode decomposition and Z-scores is developed for deployment at the edge, achieving an outlier feature recognition accuracy of 83.4%. For highly digitized converter stations or those equipped with GPU computing power, an enhanced long short-term memory-fully convolutional network-FCN (E-LSTM-FCN) network is proposed. By integrating long-term and short-term memory networks with a multihead attention mechanism, this model achieves a detection accuracy of 96.9%. Finally, ablation experiments validate both the accuracy and real-time performance of the proposed network architecture. This study bridges a critical technical gap in the acoustic monitoring of valve hall discharges and establishes a robust foundation for intelligent, real-time, and online monitoring of valve hall equipment.