The backend engine can detect and replace missing data, outliers, or biased (drifting) data with model-predicted values. Identification of missing data and outliers is elementary and is done first. We then use a physics-based Bayesian network model to validate the sensor values against the model-predicted values. When we determine drift or bias to be above a certain threshold, the model-predicted value may be used to replace the sensed value.

Some key points to note about data cleansing:

• The raw data are never discarded.
• The model-predicted value is an additional channel that may be used only carefully and when appropriate.
• The model used changes with the rig states and the model-predicted values are not always correct but overall the data do become cleaner.

Our publications provide more details on technique used.