Partial discharge detection method for switchgear based on hierarchical clustering

Accurate partial discharge (PD) diagnosis in switchgear is vital for power grid reliability. To address the challenge of discriminating PD pulses from diverse sources and noise, this paper establishes a complete workflow encompassing acquisition, denoising, and recognition. A custom measurement device coupled with bushing sensors is developed to capture raw PD signals. Subsequently, a hierarchical clustering algorithm based on equivalent time-frequency (TF) maps is proposed to effectively separate and remove four types of noisy PD pulses. Comparative analysis demonstrates that the Ward distance metric yields optimal separation performance. The purified pulses are then converted into phase-resolved partial discharge (PRPD) spectra, which are classified by a designed convolutional neural network (CNN). Validated on a dataset encompassing multiple defect types, the CNN achieved a recognition accuracy of 99.25%after pre-training. The integrated method provides a novel and effective solution for automated, high-accuracy PD detection in switchgear.