Their purpose, implementation and long-term maintenance
The purpose and benefit of soft sensors
Soft sensors or inferential sensors are implemented to make real-time predictions of a measurement that would have otherwise taken a long time to obtain. The predicted value is then used to make decisions - such as terminating the process, or making feedback control adjustments.
For example, the kappa number in the pulp and paper industry takes about half an hour to obtain from a standard laboratory test. But using readily available process measurements around the digester, such as temperature, chip rate, steam flow, alkali, and so on, we can infer the kappa number in real-time.
In some industries, such as cell cultures, laboratory measurements are sometimes available only two months later. Real-time measurements such as pH, dissolved oxygen and on line near-infrared spectra can be used to predict these quality metrics for real-time process monitoring and adjustment.
The economic pay-off can be great: improved control leads to much less off-specification product, with savings measured in millions of dollars per year for many industries.
Implementing a soft sensor
ProSensus uses multivariate methods to implement our soft sensors. This leads to the following advantages* over linear regression, neural networks and other techniques:
- our soft sensors can easily handle missing data;
- new data is assessed for validity before making a prediction - if the new data are not valid, this indicates that the process has changed away from its normal conditions, therefore any soft sensor predictions are not valid;
- realistic and valid confidence intervals can be placed on the predictions.
* A technical white paper is available that explains this further.
The following approach is followed when investigating a new soft sensor application:
- Assess the current status, and the economic benefits if the soft sensor objectives are met.
- Gather existing laboratory measurements (the value that will be predicted in the future) and historical records of on-line data (the values that will be used to make predictions)
- Build multivariate models that demonstrate the soft sensor's ability. Evaluate the soft sensor's improvement over existing systems.
- If the economic evaluation shows justification, then implement the soft sensor in parallel to the existing systems.
- Evaluate on-line performance, and complete the commissioning of the new soft sensor.
The total time for implementing such a sensor can range from several days to weeks, depending on the degree of accuracy required, and available systems and personnel.
An interesting overview of soft sensors in the pulp and paper industry is provided here from the CANMET Energy Technology Centre.
Long-term maintenance of an inferential sensor
Soft sensor maintenance is a very important issue, and is a key part of selecting a suitable inferential sensor technology.
The real-time process measurements used in the prediction come from an unsteady-state process - the process will drift and change with time. The soft sensor model must accommodate these changes.
ProSensus has developed several levels of technology that leads to long-term sensor use - from as simple as bias and scaling updates, to as sophisticated as our proprietary, L-shaped multi-block methods.
Contact us about your application
Please contact us to evaluate your soft-sensor application; pre-feasibility studies are risk free, performed at no cost.
