EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

AcousticMoEKAN: Sparse Mixture-of-KAN Experts for Time–Frequency Forecasting of OLTC Acoustic Signals

Donghyun Kim, Jimyung Kang, Yeji Hyun, Hochul Kim

IEEE Access · 2026 · 14 · 40114-40136

DOI: 10.1109/access.2026.3670058 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Donghyun Kim und 3 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Korea Institute of Energy Technology Evaluation and Planning (KETEP) and the Ministry of Trade, Industry and Energy (MOTIE) of Republic of Korea.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

On-load tap-changers (OLTCs) regulate distribution voltage and undergo frequent mechanical operations that can accelerate wear. Acoustic monitoring is non-intrusive, but forecasting abnormal behavior directly from 44.1 kHz waveforms is computationally prohibitive for multi-step prediction. We formulate OLTC acoustic prognostics as time–frequency forecasting: each tap-changing event is converted to a short-time Fourier transform (STFT) magnitude spectrogram with 513 frequency-bins, and the task is to predict future spectrogram evolution. We propose AcousticMoEKAN, an acoustic-optimized mixture of Kolmogorov–Arnold networks (KAN) that expands the expert pool from the four analytic bases used in MoEKAN to ten complementary experts and replaces dense Transformer gating with sparse Mixture-of-Block Attention (MoBA) gating for efficient top-k expert routing. Experiments on a large-scale accelerated-aging dataset comprising 195,994 tap-changing events show that AcousticMoEKAN improves accuracy averaged over the DC component and four equal-width frequency bands, outperforming linear models, convolutional neural networks (CNNs), graph neural networks (GNNs), Transformers, and KAN baselines. AcousticMoEKAN reduces average mean squared error (MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) by 20.6%/9.5%/11.6% relative to MoEKAN and by 25.1%/11.0%/12.0% relative to the best non-KAN baselines (RLinear for MSE, STID for MAE/MAPE), with consistent gains across low and high frequency bands. These results indicate that expert-augmented KAN modeling and sparse MoBA gating provide an effective framework for OLTC acoustic condition forecasting in the STFT domain.

Förderung

  • Korea Institute of Energy Technology Evaluation and Planning (KETEP) and the Ministry of Trade, Industry and Energy (MOTIE) of Republic of Korea

Zitieren

APA

Kim, D., Kang, J., Hyun, Y., & Kim, H. (2026). AcousticMoEKAN: Sparse Mixture-of-KAN Experts for Time–Frequency Forecasting of OLTC Acoustic Signals. IEEE Access, 14, 40114-40136. https://doi.org/10.1109/access.2026.3670058

Vancouver

Kim D, Kang J, Hyun Y, Kim H. AcousticMoEKAN: Sparse Mixture-of-KAN Experts for Time–Frequency Forecasting of OLTC Acoustic Signals. IEEE Access. 2026;14:40114-40136. doi:10.1109/access.2026.3670058

BibTeX
@article{Kim2026,
  author = {Kim, Donghyun and Kang, Jimyung and Hyun, Yeji and Kim, Hochul},
  title = {AcousticMoEKAN: Sparse Mixture-of-KAN Experts for Time–Frequency Forecasting of OLTC Acoustic Signals},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {40114--40136},
  doi = {10.1109/access.2026.3670058}
}
Kostenlos starten →