Anomaly Detection
Anomaly detection spots data points, events, or patterns that are very different from normal behavior. In industry, it finds equipment problems before they cause failure.
Types of Problems
- Point problems — a single data point is abnormal
- Contextual problems — a data point is abnormal in context
- Collective problems — a group of data points is abnormal together
Methods
- Statistical — deviation from statistical distribution
- Machine learning — autoencoders, isolation forests, one-class SVM
- Deep learning — LSTM autoencoders for time-series data
- Hybrid — combining multiple methods for reliability
How VLTHRLAB Uses It
VLTHRLAB uses LSTM autoencoders to detect problems in time-series data, achieving AUC up to 1.000 on 8 benchmark datasets with zero false alarms when tested on new datasets.