BROWSE
Detail
Conference Paper
경량 Trident-ResCNN 기반 Gas Leakage 및 Partial Discharge 초음파 음향 이벤트 분류 성능 분석
Cited - time in 
Share 
Share
- Issue Date
- 2026-06
- Citation
- 대한전자공학회 학술 대회 (하계) 2026, pp.1-5
- Publisher
- 대한전자공학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- This paper presents LC Trident-ResCNN, a low-complexity network for ultrasonic acoustic event classification targeting gas leakage (GL) and partial discharge (PD) under edge constraints. The architecture combines Trident-style band-splitting branches, a ResNet backbone, and depthwise separable convolutions, satisfying the DCASE Task 1 low-complexity budget (≤128 kB, ≤30 MMACs). Using 4-fold cross-validation repeated 10 times, we systematically analyze (i) frequency-bin splitting (EBFS vs. UBFS), (ii) audible-band inclusion (5–20 kHz), and (iii) knowledge distillation (KD). Extending the effective band to cover the audible range under EBFS with log-mel features improves accuracy by 1.918 pp (GL) and 1.002 pp (PD), yielding 99.34%/98.84%. Applying KD further improves accuracy by 0.317 pp (GL) and 0.354 pp (PD), for final scores of 99.425% and 99.192%, narrowing the gap to a larger teacher model.
- KSP Keywords
- Acoustic Event Classification, Cross validation(CV), Gas leakage, Knowledge Distillation, Low complexity, Partial Discharge(P.D), Teacher Model, frequency-bin