Context Presentation
Anomaly detection in real fluid distribution applications is a difficult task, especially when we seek to accurately detect different types of anomalies and possible sensor failures. Our case study is based on a real context: sensor data from the SGE (Rangueil campus management and operation service in Toulouse).
We propose an automatic pattern-based method for anomaly detection in time-series called Composition-based Decision Tree (CDT). We use a modified decision tree and Bayesian optimization to avoid manual tuning of hyper-parameters. Our method uses sequences of patterns to identify remarkable points corresponding to multiple anomalies. The compositions of patterns existing into time-series are learned through an internally generated decision tree and then simplified using Boolean algebra to produce intelligible rules.
Our approach automatically generates decision rules for anomaly detection. All our experiments were carried out on real and synthetic data. We show that our method is precise for classifying anomalies compared to other methods. It also generates rules that can be interpreted and understood by experts and analysts, who they can adjust and modify.
Keywords
Anomaly detection, Time-series, Machine learning, Classification rules
Scientific goals
• To detect different types of anomalies observed in real deployment
• To generate interpretable rules for anomaly detection
• To use learning methods for anomaly detection on static and continuous data
Contact
ines.ben-kraiem@irit.fr
