Context-Enhanced Anomaly Detection Using Deep Learning with Root Cause Characterization in Vacuum Conveying Time Series

Sammanfattning

Industrial vacuum conveying systems are critical for bulk material handling, yet current monitoring relies on threshold-based rules that cannot detect subtle anomalies. This thesis proposes a two-stage pipeline combining unsupervised deep learning for anomaly detection with interpretable machine learning for root cause characterization. In Stage 1, two reconstruction-based deep learning models (a State Space Model (SSM) (KambaAD) and an attention-based transformer (Anomaly Transformer)) are evaluated on real vacuum conveying data. Results show that including operational context (machine settings) significantly improves detection, with both models achieving Area Under Receiver Operating Characteristics (Curve) (AUROC) above 0.96 when context is available. In Stage 2, fault-specific signatures derived from SHapley Additive exPlanations (SHAP) analysis enable a rank-based characterization approach that achieves 96% top-3 accuracy, allowing operators to quickly identify likely fault types. Key contributions include: (i) a curated industrial dataset covering 30 configurations and 6 fault categories, (ii) an empirical comparison of transformer and SSM architectures for industrial Time Series Anomaly Detection (TSAD), and (iii) a complete two-stage pipeline designed for industrial deployment where interpretability is essential.