Neural computing in pharmaceutical products and process development - Computer-Aided Applications in Pharmaceutical Technology - page 123

processing system. Granule images measured by a particle image probe were
digitized by an image processing system during granulation, to continuously
calculate granule size distribution. To avoid excessive granule growth, fuzzy
logic using a linguistic algorithm employing if-then rules, with a process lag
element taken into consideration, was adopted for the control algorithm.
The neuro-fuzzy system was applied to study the effects of formulation
and processing parameters on roller compaction (Mansa et al., 2008).
Adaptive (learning) capabilities of neural networks and the linguistic
capabilities of fuzzy logic were combined in the software used ( FormRules
and INForm , Intelligensys). An MLP network was used as the 'neuro'
part of the adaptive system. Input data were generated from material
characterization studies and from investigations conducted on a
laboratory-scale roll press with side plates, where process parameters
such as roll speed, roll gap, and compaction pressure were selected as
outputs. Some of the fuzzy rules generated are:

■ IF ratio of microcrystalline cellulose to dicalcium phosphate anhydrous
is low AND roll speed is low AND compressibility is low THEN
ribbon density is low (confi dence factor is 1.0).
■ IF the roll speed is low AND the roll gap is high THEN ribbon porosity
is low (confi dence factor is 0.93).
■ IF the roll speed is medium AND the angle of wall friction is medium
THEN the average max pressure is low (confi dence factor is 0.85).

The neuro-fuzzy system was applied to analyze pellet properties (Mendyk
et al., 2010). Neural networks were used to build models correlating
input information describing quantitative and qualitative composition of
pellets, as well as technology of preparation with output variables such
as MDT and median AR. Once the sensitivity analysis was performed,
the number of inputs was reduced, since it was concluded that some of
the inputs do not affect output(s) to a greater extent. If possible, the
number of hidden nodes in neuro-fuzzy systems was reduced, in order to
obtain the simplest logical rules sets with the same predictive ability as
the original ANN.
Some of the fuzzy rules generated are:

￿

￿ ￿

■ IF the screw speed is low AND the number of holes is low AND the
spheronization speed is high AND the drying temperature is low AND
the drying time is low THEN the AR is high .
■ IF the screw speed is moderate AND the number of holes is high AND
the spheronization speed is high AND the drying temperature is high
AND the drying time is high THEN the AR is high .