RESEARCH SUMMARY
It
is imperative therefore, to device an optimal inspection policy in
conjunction with planning, scheduling and resource maintenance decisions. In
the first phase of the work, a partially observable Markov decision process
(POMDP) is solved for systems with stationary big resources and different
flow trajectories (single machine, serial flow, re-entrant flow etc). The
machines considered in this case are assumed to be very costly to be replaced
and an optimally timed preventive maintenance is required. Big manufacturing
equipments, reactors, furnaces etc fall in this category. The inspection
related to resource degradation is part of decision-making. It is shown that
under certain assumptions, the optimal decision-rule is monotone in the state
of the system. While
preventive maintenance schedule is needed in the above mentioned case,
presence of multiple small equipments (non-stationary resources) requires an
inventory of these resources to be maintained in the system. Examples include
small equipments like masks, molds, spare parts etc. People may also be
treated as resources, which need appropriate skill acquisition before getting
assigned to a task/project, and tend to exit the system depending on several
factors. When resources belonging to this category tend to break/ exit the
system, with time scales comparable with that of production, their inventory
needs to be controlled together with that of the product. This gives rise to
a special class of resource planning problem when production planning has to
be simultaneously addressed. A rolling horizon mixed integer linear program
(MILP) and an approximate Dynamic Programming algorithm are used to solve the
resulting problem and findings are compared. Finally,
the nodes of a network are considered to be resources facilitating material
or information flow, for example, food, water and
web networks. These nodes are prone to accidental or malicious attack leading
to downstream nodes getting infected. The need arises to track down the
location of the attacked node and shut down/repair the infected downstream
nodes. The decision-making around network security is formulated as POMDP and
efficient algorithms are developed for solution of the large size problem. CURICULUM VITAE Journal
Conference
PRESENTATIONS (Other
than those noted above)
|
||
|
|
||
|
On the structure of optimal policy for
machine maintenance and job inspection in systems with deteriorating equipments
and defect propagation Abstract— Random deterioration in manufacturing equipment
is traditionally modeled as Markov Decision Processes (MDP) where machine/equipment
condition is partially observed by means of observations related to resource
degradation. In systems where deterioration is reflected in increased
production of defective jobs, the inspection of operated jobs constitutes
indirect observations of the resource state.
However, economic considerations may dissuade the decision-maker to
undertake inspection of every job. This gives rise to two additional features
to the Partially Observed MDP (POMDP) (i) the observation/ inspection becomes
part of the decision-making, and (ii) the untested intermediates propagate
through the system. The latter is true for most practical manufacturing
systems with multiple operations. In this work, we consider these two
additional features by formulating the POMDP for a single machine, a re-entrant
flow station and a hybrid flow system. The large size problems thus obtained
are solved using approximate solution methods for POMDPs. It is shown that
the structural properties of the optimal value function associated with the
single machine POMDP are preserved even when the observations are a part of
the decision-making. |
||
|
Dynamic Sensor
Network Design in Chemical Plants for Fault Detection and Quality Control Abstract - Sensor network design (SND) in chemical plants is
aimed at determining optimal allocation of sensors to detect faults/errors
for process improvementand quality control purposes. This problem has been
studied in the past for maximizing observability,resolution
and reliability of the network. In case of requirements for off-line sensing
however, the need for a dynamic framework for SND arises. This coupled with
drifts associated with different processes and equipments in the plant gives
rise to a stochastic optimization problem for dynamic allocation of sensors. The
fact that not every process stream is tested coupled with the possibility of
sensor failures and erroneous sensor readings lead to partial information
about the state of the system at all times. For this reason, beliefs on
equipment health need to be maintained. In this work, we present the problem
of dynamically allocating off-line sensing capability for quality control in
a serial processing system. This problem is formulated and solved as a Partially
Observable Markov Decision Process (POMDP). Formulation as POMDP leads to
increased problem size but provides a comprehensive framework to take into
account the several trade-offs associated with the problem. |
||
|
Information Flow
Based Decomposition of Decision-Making Problems Involving Partial
Observability Abstract- Industrial planning and scheduling decisions are often
inter-dependent. For example, planning level capacity allocation decisions
affect production scheduling. While independent decision rules fail to
address the above mentioned inter-dependence, simultaneous consideration of
all interactions leads to a very large problem, which is oftentimes
computationally intractable. In this study, we demonstrate, in the context of
a machine maintenance problem for a reentrant flow system, a middle-ground
approach by recognizing paths of strong information flow and then
systematically decomposing the problem to be able to obtain a computationally
tractable problem yielding a near-optimal decision policy. In the process, we
make combined use of rigorous probability theories, approximate dynamic
programming and simulation based rules. Planning and
Scheduling in Manufacturing with Perishable Resources Abstract- The term ‘perishable resource’ stands for
a machine, equipment or tool group that breaks with time scales comparable to
that of production. This phenomenon gives rise to the need for inventory
control studies at the resource level. In certain cases, resource ageing and
eventually breaking rate is dependent on resource utilization and its
allocation to different jobs. The result is a coupling between the resource re-order
and the production scheduling decisions. Due to the difference in time scales
of the decisions at planning and scheduling levels, the size of a combined
large problem is very large. We present a hierarchical treatment to the
mentioned problem and study the information that is exchanged between the levels.
The planning problem is solved using Real Time Approximate Dynamic
Programming (RTADP) and the scheduling level decisions are made using
parameterized heuristics optimized through simulations. A case study on
perishable resource inventory control of ‘stone veneer’
manufacturing is presented to illustrate the methodology and findings. |
B.S.,
Chemical Engineering (2004), Indian 
M.S.,
Industrial and Systems Engineering (2006), Georgia Institute of Technology,
Atlanta, USA
Last Modified , LIDCUS © 2009