A Hybrid Vibration-Based Anomaly Detection System for Predictive Maintenance of Capping Equipment

This paper presents a data-driven methodology for anomaly detection in industrial capping machinery, a critical asset in high-volume manufacturing. The proposed approach employs a two-stage framework, beginning with a diagnostic phase that uses an ensemble of signal processing and statistical methods (including Power Spectral Density, Short-Time Fourier Transform, RMS Envelope Analysis, Dynamic Time Warping clustering, statistical feature analysis, and stationarity testing) to identify anomalous temporal windows by comparing real-time vibration data against a healthy operational baseline. Subsequently, a punctual anomaly detection phase within these windows leverages a voting mechanism integrating the outputs of Long Short-Term Memory (LSTM) networks, autoencoders, and Isolation Forests. Validated on industrial capping equipment under realistic production conditions through a collaboration with AROL Closure Systems, the methodology demonstrated its performance across various simulated fault conditions, including bushing wear and ring loosening. Specifically, near-perfect accuracy was achieved for drift-in- time anomalies (F1 scores of 100% for both the experiments). In transient anomaly detection, the system yielded an F1 score of 100% in the first and 98% in the second experiment. Furthermore, in challenging mixed anomaly scenarios, overall F1 scores exceeded 99%. These results underscore the effectiveness of the proposed vibration-based anomaly detection system for predictive maintenance in critical industrial machinery, offering significant potential for early fault detection, reduced downtime, and optimized maintenance scheduling.