BROWSE
Detail
Journal Article
An Encoder-agnostic Gaussian Mixture Framework for Unified Time-series Analysis in Manufacturing
- Authors
- Hyuntae Kim, Eunseo Lee, Hyun-Chul Kang, Jiyeon Son
- Issue Date
- 2025-10
- Citation
- IEEE Access, v.13, pp.2169-3536
- ISSN
- 2169-3536
- Publisher
- IEEE
- Language
- English
- Type
- Journal Article
- Abstract
- Unsupervised time-series clustering and timely anomaly detection are critical in manufacturing, where large volumes of streaming data are collected but labeled information is scarce. These tasks support both the discovery of underlying operational patterns and the early detection of abnormal behaviors – capabilities essential for reliable process monitoring. Nevertheless, deploying time-series models in manufacturing faces two major challenges: 1) many existing models rely on specific modeling choices that lack the flexibility to accommodate process-driven data shifts, and 2) diverse operational states induce heterogeneity among normal instances, limiting the effectiveness of conventional anomaly detection methods. To address these issues, we propose a modular, encoder-agnostic framework that is universally compatible with diverse architectural and learning assumptions. Our method first learns temporal representations via self-supervision and then applies Gaussian mixture modeling to structure a cluster-friendly latent space that accommodates operational diversity while enabling likelihood-based anomaly scoring. We improve robustness through numerically stable likelihood computation and practical update schemes for mixture parameters, facilitating seamless plug-and-play integration with a wide range of temporal encoders. Our method consistently outperforms strong baselines in both clustering and anomaly detection tasks, including the challenging setting of multi-class anomaly detection, as demonstrated through comprehensive experiments on benchmark time-series classification datasets. We further validate the framework’s effectiveness through evaluations on a real-world industrial dataset, highlighting the practical value of an encoder-agnostic design in manufacturing environments.
- KSP Keywords
- Abnormal behavior, Detection Method, Early detection, Gaussian Mixture Modeling, Gaussian mixture(GM), Latent space, Mixture parameters, Plug-and-Play, Process monitoring, Real-world, Self-supervision
This work is distributed under the term of Creative Commons License (CCL)
(CC BY NC ND)
(CC BY NC ND)
