Scheduling at the
company is dynamic, it happens on a rolling horizon as opposed to the typical situations
in this subject. There is no distinctive start or end point from the schedule
as jobs are continuously added to the system. As a consequence, jobs with an
availability time larger than the current time are unknown. The schedule has to
be reevaluated continuously as new jobs are added to the system. The makespan of
the schedule cannot be defined because of this endless character. The
performance of a schedule will be evaluated by looking at the leadtime of
production orders, which is defined as its completion time minus its
availability time.
Another difference in
comparison to the general job shop problem is the presence of sequence
dependent setup times. That means that the total production times are not known
exactly until the schedule is made. A production order at the shop can only be
started when the Processor is available, and an Operator is ready to start the
production.
Order sequencing is
the search for a schedule that defines start times of production orders in such
a way that certain objective functions are optimized. The objective in the
Piping production environment at the case company was to minimise average lead times
of the products. Production lines at the company have a higher variety and
lower volume than typical mass production lines. Piping section has 20
workstations, most of them contain one or two identical CNC-machines. Operators
perform a fixed setup at the start of every production order and a variable
changeover depending on the current settings of the machine and the required
settings of the next production order. The used dataset contains 660 different order
types, on average these are produced on only 2 workstations.
The Piping facility is
a flow shop; the flow of production orders between workstations is always in
the same direction. Further important characteristics are the rolling horizon
and the sequence dependent changeover times. New production orders are
continuously added to the system, there is no clearly defined start or end
point. The situation is called dynamic. Further, the changeover time of a
workstation is both dependent on the last produced product type and the next one.
The objective is to find an algorithm that will provide the operators what
production order they need to start next, in a way that minimizes overall lead times
for the production facility.
A lot of variations
exist on the traditional order sequencing problem and a lot of research has
been done about it. A simulation model was built to evaluate several priority rules.
The current sequencing method was implemented to validate the simulation and to
have a reference point to evaluate possible improvements. Currently the
operators combine a First-In-First-Out policy with a minimisation of changeovers.
In general, they select the oldest production orders and within this group they
will minimize the total amount of changeovers. The data obtained from the simulation
was then compared to data from the production facility to validate the model.
The simulation model approximated the real situation, though not with great precision.
The simulation was also validated with a second dataset to confirm its
robustness. The model allows rapid comparison of several priority rules. The
simulation also provided means to optimize small adjustments to the sequencing
algorithm. The model was made as general as possible to allow the company to
evaluate the effect of several possible changes to the production environment,
and to possibly be used for further research to optimise the production
facility.
Several rules were tested in the company and the best performing
rule was selected. The selected rule assigns to each production order the value
of its changeover time (CO) plus its average processing time per operation
(AVPRO). The next production order to be processed on a workstation, is then
the production order with the lowest value from all available production orders.
In the simulation, this rule decreased the average production leadtime per
order type with more than 20%.
The implementation was started on one workstation. The initial
results were promising; the obtained leadtime reduction of 25% was even larger
than expected from simulation, while the leadtime of a control group lightly
increased. It was expected that the leadtime of order types with a large value
for AVPRO would increase, and this behaviour was observed from the
implementation.