Instruments are essential in a process industry for continuous monitoring and control
of different process variables which are required for different purposes: meeting stringent
effluent norms, operating within safety limit, achieving desired product quality, etc. In most
cases, meeting these criteria demands accurate online continuous measurement of quality
variables. For measurement of physical process variables (temperature, pressure, flow, etc.) in an
industry, there is a large variety of accurate sensors and transducers which are quite affordable
and reliable. However, when the issue is to measure chemical or biochemical variables related
to composition (for assessment of product quality), the situation is more difficult. In most of
the cases, analysis is performed in the laboratory by collecting samples and the response time
can be in minutes or hours. Even if an online analyzer is available, the cost, precision and
reliability of such instruments are quite unsatisfactory. Moreover, some common problems
associated with hardware sensors are: time consuming maintenance, regular calibration,
age-deterioration, insufficient accuracy, long dead time and slow dynamics and large
noise and low reproducibility [1]. However, to maintain quality specifications,
instantaneous composition knowledge is necessary for implementation of an efficient control
system. Measurement limitations leading to ineffective process control may cause problems such
as low product quality, product loss, energy loss, toxic by product generation, and safety
problems. A software sensor, virtual sensor, inferential sensor or soft sensor must be considered as
a possible alternative for continuous online measurement of such variables for which
hardware sensors are not available or have less reliability and significant time delay. Soft sensor
design is an emerging area attracting huge research interests. Presently research has been done
on some industrial processes such as prediction of melt index in polymer process [2,3],
water quality parameters in sewage treatment [4], film thickness, roughness and growth rate
in chemical vapor deposition [5], crude and petroleum fraction properties [6,7], estimation
of product composition in batch distillation process [8], textural properties of snack food
[9], polyethylene terephthalate (PET) viscosity [10], quantity and composition of hot
metal from blast furnace [11] and estimation of biomass and substrate [12].
A soft sensor is an inferential model which can estimate online difficult to measure
process variables using other available easy to measure online process variables such as
temperatures, pressures, flow rates, etc. Soft sensor is a combination of two terms: (1) `software', because
the models are mostly computer programs, and (2) `sensors', because the models provide
similar information as their hardware counterparts. A typical soft sensor is composed of
three elements: the process model, the variables used by the model and an update technique.
The heart of a soft-sensor constitutes the model of a process which takes values of the
easily measurable process variables and predicts the output which is a difficult to measure
process variable thereby replacing or assisting a real physical sensor. This plant model may be a
first principle model, blackbox model or gray box model, substituting some physical sensors
and using data acquired from some other available ones.
|