Anomaly and Incident Detection
One of the standard applications of Machine Learning and Pattern Recognition is the automated detection of anomalies, incidents and failures, like it is done for diagnosis and warning systems.
With Machine Learning it is possible that normal, expected data observations are trained accordingly into a mathematical model. New data which is not fulfilling this expectation base and being different from all the previous training data, will be detected and alerted then.
Another possibility is the observation of the error from time series forecasting models like they are mentioned in time series prediction. If the prognosis does not fit the later observation, an abnormal situation is detected and an alert can be triggered.
If new data has been classified as abnormal but should not in future, the data can be included into the expectation database simply by adoptively training it to the model. Hereby the anomaly detection can be implemented in an adaptive way to learn new situations accordingly.
That way one for example can
- evaluate traffic flow and detect accidents or other abnormal situations,
- implement pollutant warning systems for environmental protection,
- identify damaged or conspicuous data in large data sets,
- identify failing test immediatety on the fly,
- implement diagnosis systems for automated and learning machine and plant monitoring,