Some of our customers have used NeuroShell to build process control applications. The following tips will assist you in building similar applications.
The typical process control application involves the setting of a number of process controls based upon a number of environmental variables and a quality factor.
The easiest way to visualize how to set up a process control application is by hypothetical example. Suppose you are in charge of a waste water decontamination plant where you have several knobs or switches which control how the waste water is processed (the process we want to control). Let's call these knobs or switches K1, K2, ...Kn. They will control things like pressure, heat, churning action, etc.
There will also be a number of environmental factors E1, E2, ...Em which affect how the waste water needs to be processed. The Ej will be factors like the temperature of the water, the percentage volume of solid matter, the time of day, and others.
Finally, the quality factor is some objective measure of how good the effluent from the plant is. A quality factor of 1 may mean that it's good enough to drink, and 0 may mean it's virtually untreated.
An immediate NeuroShell 2 setup that comes to mind follows: Here the Ej and Ki are inputs to the network that is used to predict a Quality Factor:
E1, E2, ...Em, K1, K2, ...Kn.
The logs of the treatment plant will hopefully be complete enough that there are many months or years of data showing how well the effluent was treated with different knob settings, and hopefully there are some mediocre operators who have not always been able to produce high quality output. That way we will have enough different data to build a network capable of handling a wide range of variables.
Predicting Knob Settings
The network described above will predict how good the effluent will be under given environmental conditions with given knob settings. However, that does not really do what we want, since we need to know how to set the knobs.
Consider, then, this alternate arrangement:
E1, E2, ...Em, Quality Factor
K1, K2, ...Kn
Here we have made the desired quality factor an input along with the current environmental factors. The network sets the process control knobs for you.
As with most networks, the more sample patterns you have, the better. As previously stated, you will need a number of patterns where the effluent was poor, even if you have to generate them by knowingly setting the knobs wrong to see how the quality varies.