If you have built predictive models with any number of techniques – advanced pattern recognition is popular in power generation – you know the challenge of selecting training data. One tip that is simple but will make this activity easier (and your models better) is described here.
Instead of looking at the upper and lower limits (range) of the data and assessing abnormal points, take a more pragmatic approach. Gather a list of alert limits – operator warnings and alarms, engineering limits, etc. Use this list of limits along with the time-series data to view trends toward the limits. Trends that are approaching or are outside of the limits should be removed from the training data.
For example, see the picture for a time-series plot of bearing temperature in degrees Fahrenheit versus time with lower and upper Limits shown in yellow and red, respectively. The bearing temperature exceeds a lower limit during the time-frame highlighted in yellow. These records should be removed from the training data set (excluded or deleted, whichever your software allows). Note, it is often a good practice to remove records that are approaching the lower limit. The reason for this is that the predictive models should detect abnormal changes prior to approaching an already known limit.
Please write me at aaron.hussey@int-analytics with comments or questions.