Genetic_Programming_Theory_and_Practice_XIII

Using GP for Data Science 123

Process Input Types

Temperature

Pressure

Vibration

Chemical Composition

Percentage mix

Output of

upstream process

Physics based model

Simulation

Back-calculation

Inferred or

Calculated Inputs

Inputs from

Low Level

Sensors

“State” of

Input

Derived Inputs

Machine/

Manufacturing

Plant/Process

“product”

Fig. 1 Different input types to the operations optimization problem

and requires significant data processing and cleaning prior to use. Therefore, the
problem at a high-level is one of using real-time sensors and control capability
to understand and improve operations. Automated methods that clean up data or
estimate non-measurable attributes help in providing an accurate, real-time view of
the whole system. In Kordon and Smits ( 2001 ), the authors use GP to create soft
sensors, or virtual sensors, that augment more expense sensors.

3.1 The Data Science Challenge at Hand

Our Data Science challenge could be stated as follows: Given historical data of an
operation, is it possible to use data and analytic methods to create accurate sensor
estimators, for those time instants when the sensor is offline? A sensor often changes
from being online (available) or offline (unavailable). A secondary challenge was to
identify instances (time periods) when the sensor of interest is drifting away from its
ideal accuracy level or from its prior relationship with other system sensors. Solving
these challenges would give an operations managers a consistent, real-time stream
of data that characterizes their operation, enabling accurate and timely optimization
decisions.
We defined our problem as having sensorss 1 :::sN 1 as available, and sensorsN
as partially available: meaning that for portions of time sensorsNis online, but it
frequently goes offline due to other activities or faults. Our goal was to determine
whether we could build an accurate model using data for sensorss 1 :::sN 1 and