The basic thrust in this area is the development of approaches that
combine control technology with optimization methods and
applied artificial intelligence techniques to create systems that
can (to the extent realistically possible) optimize their own
performance and, in the process, learn how to refine their models
and knowledge bases to continually improve their efficacy. At the
core of such a system, one has the plant or process
including the usual instrumentation and control loops designed by
conventional means for conventional ends (e.g., set-point regulation).
Most industrial processes are built in this fashion and operated
according to set-point time-histories or ``recipes'' that define
the standard approach for producing the desired product. The
problems with using off-the-shelf recipes in this way are indeed
serious:
- variations in raw materials and operating conditions may result in
out-of-spec products and wastage, and
- generating new or modified set-point recipes to meet differing specifications
and constraints is a time- and resource-consuming process.
To meet these challenges, our ISOC framework appends to
the process a model for cost in the very broad sense, including labour,
energy and raw materials, expenses associated with failing to
meet environmental regulations, costs associated with quality control,
and the like. Together, the process and cost model represent the system
to be controlled in an intelligent and optimal manner. As shown below,
this approach is comprised of the following elements:
-
an intelligent supervisor that uses a combination of AI methods for
defining a suitable model-based optimization problem plus optimization routines for
solving it; this creates the optimized recipes needed for plant operation;
-
a model refinement module that continually monitors the process and
cost variables to tune the model parameters to increase fidelity
and track changes; and
-
a historical database of past operations data that can serve as the basis for
case-based reasoning to pick the nearest match(es) to the present run
specification and thereby establish a feasible starting point for the
optimization process.
These modules form a model-based system that learns from experience, in
the sense that the models are continually updated to reflect the possibly
changing behaviour of the process and cost factors as accurately as possible.
The optimization procedures are guided by an increasing body of knowledge
about recipes that were successful and those that were not.