ETRI-Knowledge Sharing Plaform

KOREAN
논문 검색
Type SCI
Year ~ Keyword

Detail

Conference Paper Beyond Clean Evaluation: Stress-Testing Time-Series Backbones under Manufacturing Sensor Faults
Cited - time in scopus Share share facebook twitter linkedin kakaostory
Authors
Jeongeun Kim, HongYeon Yu, Gihyeon Min, Dong Hoon Son, Yun ji Ban, Seihyoung Lee
Issue Date
2026-10
Citation
International Conference on Information and Communication Technology Convergence (ICTC) 2026, pp.1-4
Publisher
IEEE
Language
English
Type
Conference Paper
Abstract
Manufacturing time-series systems are often selected on clean data although deployed sensors may be noisy, biased, missing, or unavailable. We compare frozen-core PatchTST, TimesNet, and iTransformer on CWRU, MachineVision, and Hydraulic data under reference (Base) and manufacturing-pretrained (Mfg-PT) weight states. Five downstream seeds are crossed with five shared realizations of Gaussian noise, persistent offset, synchronous masking, and mixed fault proxies; Hydraulic additionally receives asynchronous masking and sustained channel loss. We report clean and equal-weight stress macro-F1, worst-condition mean degradation, paired Mfg-PT–Base stress-mean differences, and validation-selected test regret. Across the fixed pool, no system was uniformly best. On CWRU/Base, iTransformer had the highest clean mean (0.626), whereas TimesNet had the highest stress mean (0.520). Mfg-PT increased the stress-mean point estimate in only two of nine system–dataset pairs; four paired 95% confidence intervals were below zero and three included zero. The largest mean validation-selected stress regret was 4.82 [3.14, 6.50] pp on MachineVision/Mfg-PT. Clean performance, fault robustness, and selection loss therefore warrant separate evaluation.
Keyword
manufacturing applications, sensor faults, time- series classification, robustness, self-supervised pretraining
KSP Keywords
Channel Loss, Gaussian Noise, Mixed fault, Series systems, Time series, confidence interval, point estimate, sensor faults