Summary
Anomaly detection automatically identifies deviations from normal patterns in production data (time-series spikes, drifts, dead values; process parameter excursions, quality drops, pace anomalies). Root cause analysis then traces detected anomalies back to their fundamental causes. The pipeline integrates statistical methods (3-sigma rules, SPC control charts, IQR), machine learning (Isolation Forest, LOF, One-Class SVM), and deep learning (AutoEncoder, LSTM-AE, Transformer). Root cause analysis leverages knowledge graphs of factory entity relationships and causal inference methods (Granger causality, PC algorithm, do-calculus).
Key Claims
- Method selection follows a decision tree: with labels use supervised classifiers; without labels, low-dimensional data suits statistical methods, while high-dimensional data with few samples favors One-Class SVM and many samples favor Isolation Forest or AutoEncoder.
- Knowledge graph-driven root cause analysis builds entity relationships (equipment, process, material, supplier) and traces anomalies along causal chains to identify the most probable root cause node.
- The end-to-end system uses Kafka and Flink for real-time anomaly detection, with offline training of root cause graphs and causal models, feeding alerts and dashboards at the application layer.
Connections
- DataWarehouse — anomaly detection integrates multi-source data (SCADA/MES/QMS) from the warehouse with time-aligned timestamps
- DataGovernance — consistent entity naming, standardized fault codes, and clean timestamp alignment are prerequisites
- RAG — root cause analysis could be enhanced by RAG systems that retrieve historical incident reports matching current anomaly patterns