•  
  •  
 

Keywords

converter valve; voiceprint; feature extraction; denoising auto-encoder; anomaly detection

Abstract

The converter valve, as a key equipment in high-voltage direct current transmission converter stations, may experience an abnormal state when operated for a long time. To address the problem that non-intrinsic background noise in the valve hall may cause voiceprint anomalies and unstable anti-noise performance of the converter valve, an anomaly detection method for the converter valve based on voiceprint signal filter banks (Fbank) features and an improved denoising auto-encoder (IDAE) is proposed. Firstly, the voiceprint data generated during the operation of the converter valve is collected; the Fbank features of the voiceprint data are extracted; the samples containing temporal information through sliding window (SW) processing are obtained. Then, a denoising auto-encoder (DAE) based on environmental noise and dual channels is constructed to train normal samples, and multiple feature thresholds are calculated through fused directional distance (FDD) reconstruction error. Finally, speech, white noise, and industrial background noise with different signal-to-noise ratios are added to the test data for comprehensive performance evaluation. The experimental results show that compared with other anomaly detection models, the proposed method has better performance and stronger noise resistance.

DOI

10.19781/j.issn.1673-9140.2025.06.017

First Page

175

Last Page

183

Download

Share

COinS