EUVIMEDEUVIMED entdecken
Crossref

Anomaly Pattern Mining for Predictive Maintenance of Industrial Machinery Using Sensor Time-Series Data

Sivakumar S, Gayathri T, Praveena S

PatternIQ Mining · 2026

DOI: 10.70023/p3202 ↗

Kurz gefasst

Fachartikel in der Zeitschrift PatternIQ Mining, 2026. Verfasst von Sivakumar S und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Industrial machinery continuously generates multivariate sensor signals that contain important information about operational states, degradation trends, and impending failures. However, conventional predictive maintenance techniques based on fixed thresholds or isolated anomaly detection often struggle to capturez nonlinear sensor interactions, long-range temporal dependencies, and evolving anomaly patterns. To address these limitations, this study proposes APM-TCN (Adaptive Anomaly Pattern Mining with Temporal Convolutional Networks), a predictive maintenance framework for discovering abnormal operating patterns from sensor timeseries data. The framework integrates Variational Autoencoder (VAE)-based latent representation learning, Temporal Convolutional Networks (TCN), self-attention, adaptive anomaly scoring, and DBSCAN (Density-Based Spatial Clustering of Applications with Noise)-based temporal pattern clustering. Initially, the VAE normalizes noisy sensor observations and transforms them into compact latent representations. The TCN subsequently learns short- and long-range temporal dependencies, while self-attention emphasizes sensor characteristics associated with progressive equipment degradation. An adaptive anomaly score combines reconstruction error, temporal prediction error, and attention-derived information to identify deviations from normal operating behavior. DBSCAN then groups related anomaly observations to reveal recurring degradation signatures, which are subsequently utilized for failure-risk prediction. Experimental evaluation demonstrates 97.2% anomaly detection accuracy, 95.8% precision, 96.4% recall, and 96.1% F1- score, together with 92.6% early-failure prediction accuracy. APM-TCN further achieves a 31.5% reduction in false alarms and a 39.2% improvement in early-warning capability compared with conventional monitoring approaches. The findings demonstrate that adaptive temporal anomaly pattern mining can provide reliable early indications of machinery degradation, supporting proactive maintenance planning, reducing unplanned downtime, and improving industrial equipment reliability.

Zitieren

APA

S, S., T, G., & S, P. (2026). Anomaly Pattern Mining for Predictive Maintenance of Industrial Machinery Using Sensor Time-Series Data. PatternIQ Mining. https://doi.org/10.70023/p3202

Vancouver

S S, T G, S P. Anomaly Pattern Mining for Predictive Maintenance of Industrial Machinery Using Sensor Time-Series Data. PatternIQ Mining. 2026. doi:10.70023/p3202

BibTeX
@article{S2026,
  author = {S, Sivakumar and T, Gayathri and S, Praveena},
  title = {Anomaly Pattern Mining for Predictive Maintenance of Industrial Machinery Using Sensor Time-Series Data},
  journal = {PatternIQ Mining},
  year = {2026},
  doi = {10.70023/p3202}
}
Kostenlos starten →