Temporal Convolutional Networks for Anomaly Detection in High-Speed Textile Yarn Monitoring

In high-speed textile manufacturing, yarn motion must be precisely controlled to avoid defects and machine faults, yet many yarn anomalies are short-lived and difficult to detect using conventional sensing systems. This thesis proposes a data-driven anomaly detection framework based on high-frequency, camera-derived yarn position measurements acquired from an industrial monitoring system. A complete processing pipeline is developed, including signal reconstruction, filtering, visualization, and phase-based motion modeling to provide interpretable representations of yarn dynamics. A labeled dataset is constructed through conservative frame-level annotation of anomalous intervals under real production conditions, and signals are transformed into structured temporal windows using engineered features capturing positional, dynamic, and cross-channel characteristics. An anomaly detection model based on Temporal Convolutional Networks is designed to learn multi-scale temporal patterns from these sequences while maintaining computational efficiency. Experimental results demonstrate high detection precision and robustness on a highly imbalanced industrial dataset, highlighting the suitability of temporal convolutional models for real-time anomaly detection in textile yarn production.