Pattern Recognition & Decision-Making in Railroad Maintenance
Todd Whittman
Master of Sciences in Mathematics (Industrial and Applied), August 1999
Abstract
To ensure the safe operation of freight railroads and a comfortable ride aboard
commuter lines, the tracks must be maintained with a smooth, rounded rail head.
Even small deformations in the rail head resulting from excessive weight and
use are capable of derailing trains. The world's leading railroad maintenance
organization, Loram Maintenance of Way, Inc., has developed the C21 Rail Grinder,
a multiple car train outfitted with a number of large grinding stones capable
of cutting iron. Each of these stones has three axes of motion and can be set
from the cab of the train to form grinding patterns. A laser imaging system
on the front of the C21 obtains a cross-sectional view of the rail and displays
the image on the computer inside the cab. The goal of the C21's operator is
to determine the proper grinding patterns and train speeds that will mold the
rail head into an ideal shape.
Loram currently owns several C21 Rail Grinders operating on four continents.
At the time of this writing, the pattern selection is made by the C21 operator
based entirely on his/her knowledge and experience. Loram's engineers determined
that the operator was selecting from only a few different patterns and making
several passes on each section of track at low speeds to slowly grind the rail
into a smooth shape. The Research and Development department was given the task
of developing computer algorithms that would efficiently determine the appropriate
combination of grinding patterns and track speeds.
A process for the C21 was developed that would be executed on the left and
right tracks independently. After the image of the actual rail head is obtained,
the first step is to identify the rail shape from among a predetermined set
of representative rail profiles. This can be done very quickly by calculating
several geometric invariants, or moments, of the rail profile. Second, the desired
extent and locations of grinding need to be determined, which can be expressed
with a metal removal (MR) curve. To accomplish this, the actual rail profile
is superimposed with the ideal profile and the difference between the two profiles
is calculated using cubic spline interpolation. Third, the grinding patterns
and track speeds are selected using previously obtained data on the representative
rail profiles. Several algorithms were tested for this purpose, including enumerative
searches, greedy algorithms, and genetic algorithms.
Research supported by the Minnesota Center for Industrial
Mathematics (MCIM)
