### PhD Research Synopsis [Download as PDF]

In the past few years, there has been significant technological advancements in different areas of process analysis and optimization. Examples of processes include manufacturing processes, such as assembly lines, and supply chains. These processes often involve physical or virtual inventories of products, parts and materials that are used to anticipate uncertainties on supply or throughputs of machines. Over time, the state of the machines, inventories and the whole process changes until process completion. Such processes are described as Buffered Temporal Flow Processes (BTFP). BTFP are particularly prevalent on discrete manufacturing such as in automotive, furniture, smartphones, airplanes and toys.

Due to increased global competition, manufacturing companies look toward ways to reduce their cost and increase efficiency of operations. Thus, there is a greater need for analysis and optimization of the operation results on the manufacturing floor while taking into account sustainability metrics. Often, the metrics in the manufacturing models contain random noise, which makes them stochastic. Also, most real world manufacturing models are non-linear, i.e., they may have non-linear objective and/or constraints. To support analysis and optimization of such models, there is a need to accurately model machines, inventories, processes etc., and then perform analysis on these models that can aid the project managers of a manufacturing floor in decision guidance.

The models for the discrete manufacturing floors need to capture: (a) control variables (deterministic or stochastic); (b) metrics of machines (such as cost, energy consumption, and emission) as a linear, piecewise-linear or non-linear function of these control variables; (c) process routing that describes the flow of materials through the manufacturing floor; and (d) intermediate material storage and distribution (work-in-progress) for inventories. Using these models, it is desirable to allow manufacturing and process operators to perform a variety of analysis and optimization tasks efficiently including what-if prediction and optimization. Research efforts to model and analyze BTFP have fallen short to: (a) provide a reusable representation of discrete manufacturing floors that is composable, easy-to-use and flexible, (b) perform analysis such as optimization and decision analysis on stochastic and non-linear real-world models efficiently, and (c) provide a standard interface to store and reuse manufacturing models and perform analysis such as computation, perdition, optimization and what-if analysis on these models through this interface. My research effort tries to fill in this research gap.

In order to model and analyze BTFP, I proposed the temporal Manufacturing Query Language (tMQL) in [1]. tMQL allows for modular composition, manipulation, what-if analysis and optimization of BTFP processes, machines and work-in-progress inventories. tMQL is flexible, extensible, reusable and easy to use. Because tMQL allows for process refinement, complex manufacturing processes can be easily modeled. The tMQL framework also contains the concept of a knowledge base where these components (both initial models and query results) can be stored and reused. The language I proposed in tMQL divides the manufacturing floor components into processes, inventories and flows. The processes map to the atomic machines or a larger composed manufacturing floor. Composed processes can encapsulate other processes, inventories and flows. It is possible to represent storage, distribution, and flow of items through the manufacturing floor at different times across the time horizon using tMQL. This allows for capturing the entire mathematical model of a discrete manufacturing floor in a modular and extensible way that will be reused for different types of queries such as prediction, simulation, optimization and learning.

In order to facilitate a standard interface to represent complex models on the manufacturing floor, a decision guidance analytics framework is proposed in [2] as well as the organization of a reusable Knowledge Base (KB) in [6]. The reusable KB will have three libraries: atomic models, composite models, and analytical views. The KB will contain specialized algorithms that automatically translate the high level, uniform representation of the manufacturing models into low level specialized models as required by each of the underlying tools. Additionally, the analytical views will contain algorithms that can handle the complexities of the manufacturing models. The proposed models can be developed in an open-source standard query and processing language such as JSONiq. The atomic and composite models can be used and reused directly in prediction, computation, optimization and learning queries.

Using a tMQL analytical model of a discrete manufacturing floor, the processes operator may want to perform computation, simulation, prediction, optimization, and learning operations. In [1], [3] and [5], I make some efforts towards building algorithms for BTFP. The syntax and semantics of the computation and deterministic optimization queries of a BTFP modeled in tMQL is given in [1] and [5]. The compute query will simply reduce the expressions that are defined in terms of other instantiated parameters and expressions to produce a fully grounded tMQL component. The optimize query uses the Optimization Programming Language (OPL), a Mixed-integer linear programming (MILP) solver to find the machine settings so as to minimize/maximize the objective subject to all constraints in the model being satisfied. These computed and optimized models can be stored in the knowledge base for future use. In order to solve the problem of stochastic optimization in BTFP where the task is to find the machine setting so as to minimize/maximize the expected value of the objective subject to the probability of constraint satisfaction being above some threshold, I propose an Iterative Heuristic Optimization Simulation (IHOS) algorithm in [3]. This algorithm uses a deterministic approximation approach where first, the machine settings are found in a deterministic environment and then heuristics are used to find if these machine settings would satisfy a desired confidence level in the stochastic setting. IHOS is compared with four popular simulation-based optimization algorithms in an initial experimental study. The experimental study demonstrates that IHOS significantly outperforms the other algorithms in terms of optimality of results and convergence time.

BTFP type manufacturing floor also needs to be resilient to failures and changes on the manufacturing floor. Initial efforts are made in [4] to model a steady state BTFP as a (a) a process model to represent machines, part inventories, and the flow of parts through machines in a discrete manufacturing floor; (b) a predictive queuing network model to support the analysis and planning phases; and (c) optimization models to support the planning phase. These efforts combine models of different nature in a seamless manner. These models can be used to predict manufacturing time and the energy consumed by the manufacturing process, as well as finding the machine settings that minimize the energy consumed or the manufacturing time subject to a variety of constraints. The main goal of this effort is to make the manufacturing floor fault-tolerant and adapt to the dynamic changes in requirements easily and efficiently.

To conclude, the intended contributions of my research are:

- Provide a reusable representation of discrete manufacturing floors that is composable, easy-to- use and flexible
- Perform analysis on deterministic/stochastic, piecewise-linear/non-linear and combinatorial real- world discrete manufacturing floor models efficiently
- Provide a manufacturing floor model that is resilient to failures
- Develop a standard hierarchical organization for data representation, manipulation and querying

with the help of JSON data format and JSONiq query language - A case study to model a real-world discrete manufacturing floor and perform optimization and

what-if analysis on this model

References

**M. Krishnamoorthy**, A. Brodsky, and D. A. Menascé. Temporal manufacturing query language (tMQL) for domain specific composition, what-if analysis, and optimization of manufacturing processes with inventories. Technical Report GMU-CS-TR-2014-3, Department of Computer Science, GMU, Fairfax, VA, USA, 2014 (was presented at the INFORMS Computing Society 2015 proceedings, it did not appear in proceedings).- A. Brodsky,
**M. Krishamoorthy**, D. Menascé, G. Shao, S.Rachuri. Toward Smart Manufacturing Using Decision Analytics at IEEE International Conference on Big Data (Big Data), 27-30 Oct. 2014, pp. 967-977. **M. Krishnamoorthy**, A. Brodsky, D. Menascé. Optimizing Stochastic Temporal Manufacturing Processes with Inventories: An Efficient Heuristic Algorithm Based on Deterministic Approximations at Operations Research and Computing: Algorithms and Software for Analytics, Proc. INFORMS Computing Society Conf., Richmond, VA, 11-13 Jan. 2015, pp. 30-46.- D. Menascé,
**M. Krishnamoorthy**, A. Brodsky. Autonomic Smart Manufacturing. Accepted at the Journal of Decision Systems, eds. I. Linden, J. Linden, S. Liu, Special Issue on Integrated Decision Support Systems, June 2015. **M. Krishnamoorthy**, A. Brodsky, and D. A. Menascé. Modular Modeling & Optimization of Temporal Manufacturing Processes with Inventories. Hawaii International Conference on System Sciences (HICSS-49) 2016 proceedings, Kauai, HI. 5-8 Jan. 2016.- A. Brodsky, G. Shao,
**M. Krishnamoorthy**, A. Narayanan, D Menascé, and R. Ak. Analysis and Optimization in Smart Manufacturing based on a Reusable Knowledge Base for Process Performance Models. IEEE International Conference on Big Data (Big Data) 2015, Santa Clara, CA, 29 Oct.-1 Nov. 2015.