Soft sensors (on-line neural nets) are spreading across the process industries at a staggering three per day! They’re in industries like refining, petrochem, polymers, pulp and paper, mining, food, pharmaceuticals, fine chemicals, semiconductors and textiles.
Why? Because they produce real live measurements, equivalent to inferred measurements – they’re virtual analysers. They also act as predictors, generating an analysis output – for now, or some time in the future. This is popular with operators who are accustomed to feedback from intermittent lab analysers, or devices with built-in delay!
Behind soft sensors
Soft sensors comprise two parts: a prediction model and sensor validation component. The former is a three-layer neural net. Inputs to the prediction model come from the sensor validation net which detects failed instruments and automatically generates reconciled measurements using a two or three layer net.
Sensor validation takes advantage of the redundancy in process instrumentation. When it’s trained, it learns the physical relationships between measurements. So if its result varies from its measured value by more than a tolerance, it’s assumed that the instrument has failed. Experience shows that about 25% of instruments fail, yet the sensor validation network still produces good artificial measurements.
The down side is that they involve large data sets. Lab data occurs at a different frequency to field data, so construction of a data set is time consuming. Process data may come from a DCS historian or an information system, or a SCADA package historian. It may be made up of sets of data that were periodically downloaded from a DCS historian. Lab data is time-stamped.
Collection amounts to gathering all this data for a representative period – including periods of steady-state. Pre-processing involves an engineer examining the data file, removing bad data, such as process upsets or shutdowns, filtering it and interpolating through missing data.
Modelling then trains the neural network using the pre-processed data. For modelling, the data is divided into three sets: training, test and validation. The training set is the set over which the back propagation algorithm is exercised. The test set is used periodically to determine the accuracy of the model.
When training and test error are minimised, the neural net parameters are saved and become the model. This is then verified by executing it over the verification set and noting how well the soft sensor output matches reality.
During the analysis step, inputs that do not contribute to the model can be removed and the model step repeated. By this procedure complex models having hundreds of inputs can be pared down to simpler, more meaningful models.
Soft sensors in action
Four software sensors were installed in 1994 by Marathon Oil in Texas City refinery. These were implemented on a crude tower and infer kerosene flash point, distillate flash point, atmospheric gas oil % boiled at 450srF, and atmospheric reduced crude % boiled at 500srF
Several software sensors were installed at Albemarle’s alpha olefins plant in Feluy, Belgium. These were based on lab samples taken every 12 hours for streams throughout the process phases. The largest model made by Albemarle had 320 points and 10,000 rows of data!
Polymer melt index is a very popular soft sensor application, probably because very accurate models result – and the alternative of a hardware analyser is expensive.
Two software sensor applications were completed in 1993 for a liner board machine at packaging Corporation of America’s Tomahawk, Wisconsin mill. The software sensor predicted Corrugating Medium Test, a strength and porosity measure for cardboard, based on 16 inputs.
Finally, two pH software sensors were installed in the First Miss Gold Getchell mine, north east of Winnemucca, Nevada in 1994. The process was the neutralisation circuit where limestone is used to increase the pH of slurry so that it can be purified into gold.
* The Author is with Pavilion Technologies. This feature is based on a paper given at ISA 96.