Toward Automated Anomaly Detection and Categorization in Polymer Fiber Production

Ensuring product quality in synthetic fiber manufacturing is critical to reducing defects and minimizing operational disruptions. This study evaluates three anomaly detection methods-rolling median, averaging-based signal fusion, and Principal Component Analysis (PCA)-based dimensionality reduction-applied to multivariate time-series data from fiber sensors monitoring properties such as magnitude, phase, node quality, and node count. Each method processes segmented 30 second windows and identifies anomalies using z-scores. Among the techniques, PCA-based fusion demonstrates superior performance by integrating multivariate signals, reducing noise, and improving detection accuracy. This approach offers a scalable, automated solution for fault detection, addressing the limitations of traditional heuristic-based methods. These findings underscore PCA's potential to enhance quality control in fiber manufacturing, paving the way for more reliable anomaly detection in industrial applications.