Three anomaly detection methods based on a comparison of signal values
with predictions from an autoregressive model are presented. These me
thods are: the extremes method, the chi(2) method and the sequential p
robability ratio test. The methods are used to detect a change of the
standard deviation of the residual noise obtained from applying an aut
oregressive model. They are fast and can be used in on-line applicatio
ns. For each method some important anomaly detection parameters are de
termined by calculation or simulation. These parameters are: the false
alarm rate, the average time to alarm and - being of minor importance
- the alarm failure rate. Each method is optimized with respect to th
e average time to alarm for a given value of the false alarm rate. The
methods are compared with each other, resulting in the sequential pro
bability ratio test being clearly superior.