Paper, We Propose an Architectural Design Performance Models and Domain-Specific Dashboards
Total Page：16
File Type：pdf, Size：1020Kb
Int J Adv Manuf Technol DOI 10.1007/s00170-016-8761-7 ORIGINAL ARTICLE Analysis and optimization based on reusable knowledge base of process performance models Alexander Brodsky1 & Guodong Shao2 & Mohan Krishnamoorthy1 & Anantha Narayanan3 & Daniel Menascé1 & Ronay Ak2 Received: 13 January 2016 /Accepted: 8 April 2016 # Springer-Verlag London (outside the USA) 2016 Abstract In this paper, we propose an architectural design performance models and domain-specific dashboards. and software framework for fast development of descriptive, Furthermore, we illustrate the use of the proposed architecture diagnostic, predictive, and prescriptive analytics solutions for and framework by prototyping a decision support system for dynamic production processes. The proposed architecture and process engineers. The decision support system allows users framework will support the storage of modular, extensible, to hierarchically compose and optimize dynamic production and reusable knowledge base (KB) of process performance processes via a graphical user interface. models. The approach requires developing automated methods that can translate the high-level models in the reus- Keywords Smart manufacturing . Data analytics . Domain able KB into low-level specialized models required by a vari- specific user interface . Optimization . Reusable knowledge ety of underlying analysis tools, including data manipulation, base . Process performance models optimization, statistical learning, estimation, and simulation. We also propose an organization and key structure for the reusable KB, composed of atomic and composite process 1 Introduction Smart manufacturing (SM) requires the collaboration of ad- * Guodong Shao vanced manufacturing capabilities and digital technologies to [email protected] create highly customizable products faster, cheaper, and greener. According to [1], “Next-generation software and Alexander Brodsky computing architectures are needed to effectively mine data [email protected] and use it to solve complex problems and enable decision- Mohan Krishnamoorthy making based on a wide range of technical and business pa- [email protected] rameters.” These software and computing architectures need Anantha Narayanan capabilities to support the development of analysis and opti- [email protected] mization solutions. These capabilities need to be designed for Daniel Menascé multiple operational levels, including manufacturing units, [email protected] cells, production lines, factories, and supply chains [2]. Ronay Ak The required analysis and optimization capabilities can be [email protected] broadly classified as descriptive (what happened?) [3, 4], di- agnostic (why did it happen?) [5, 6], predictive (what will 1 Department of Computer Science, George Mason University, happen?) [7, 8], and prescriptive (how can we make it hap- Fairfax, VA, USA pen?) analytics [9–11]. However, the current manufacturing 2 Engineering Laboratory, National Institute of Standards and practice is that analysis and optimization solutions are typical- Technology (NIST), Gaithersburg, MD, USA ly implemented from scratch, following a linear methodology. 3 Department of Mechanical Engineering, University of Maryland, This leads to high-cost and long-duration development and College Park, MD, USA results in models and algorithms that are difficult to modify, Int J Adv Manuf Technol extend, and reuse. A key contributor to these deficiencies is Corporation CPLEX mixed integer linear programming the diversity of computational tools, each designed for a dif- (MILP) solver, as described in [12]. The stochastic optimiza- ferent task such as data manipulation, statistical learning, data tion is implemented by a heuristic algorithm from [13]based mining, optimization, and simulation. Because of this diversi- on a series of deterministic approximations, which significant- ty, modeling using computational tools typically requires the ly outperforms existing algorithms based on stochastic simu- use of specialized low-level mathematical abstractions and lation. The graphical domain-specific modeling environment languages. As a result, the same manufacturing knowledge is implemented using the generic modeling environment is often modeled multiple times using different specialized (GME) [14]. abstractions, instead of being modeled only once using a uni- The paper is organized as follows. Section 2 discusses the form abstraction. Furthermore, the modeling expertise re- needs and challenges encountered in implementing analysis quired for the low-level abstractions and languages is typically and optimization solutions. Section 3 provides the design of not within the realm of knowledge of manufacturing users the architecture and framework that is based on reusable KB such as operators and process engineers. of process performance models. Section 4 extends the previ- Addressing the described limitations of current practice is ous section by exemplifying the reusable KB. Section 5 intro- the focus of this paper. More specifically, the contributions of duces the prototype for the domain-specific SM decision sup- this paper are as follows. First, we propose an architectural port system. Section 6 discusses the implementation architec- design and framework for fast development of software solu- ture for the prototype. Section 7 gives a more detailed discus- tions for descriptive, diagnostic, predictive, and prescriptive sion on related work and its limitations that we address in the analytics of dynamic production processes. The architecture paper. Finally, Section 8 concludes and discusses future work. adopts (1) the top layer of domain-specific modeling and an- alytics’ graphical user interface (GUI) and (2) the low-level layer of computational tools. The uniqueness and novelty of 2 Required analysis and optimization capabilities the proposed architectural design and framework is its middleware layer, which is based on a reusable, modular, To discuss the required analysis and optimization capabilities and extensible knowledge base (KB) of process performance in more detail, we use the diagram of the Tesla car manufactur- models. Reusability of modular KB models could lead to con- ing process example as depicted in Fig. 1. Aluminum coils are siderable reduction in the development cost, time, and the the input of the manufacturing process and are fed into two required level of expertise. The key technical challenge lies uncoiling machines that work in parallel to flatten the coils in the development of a middleware analytics engine. This into aluminum plates. The plates are then sent to four different engine comprises algorithms and automatic methods that cutting machines to prepare for the four parts of a car: the left translate high-level uniform representations of performance side, the underbody, the front, and the right side. After being models in the reusable KB into low-level specialized models cut, the aluminum plates are sent to die press machines after required by each of the aforementioned underlying tools. which they will be reinforced and welded. After assembly, the Second, we propose the organization and the key structure finished body is then washed, coated, and painted before the of the reusable KB, which consists of (1) an extensible library final operations are performed to produce a car. of atomic process performance models of unit manufacturing Different analysis and optimization capabilities are re- processes, (2) a library of composite process performance quired to analyze the performance of the production line and models, which can be constructed from the atomic process to achieve SM goals. These capabilities can be classified as performance models using a GUI, and (3) a library of analyt- descriptive, diagnostic, predictive, and prescriptive/ ical views and dashboards designed for specific types of anal- optimization analytics. ysis for domain-specific users. Descriptive capabilities are needed to create a temporal Third, to illustrate the use of the proposed design and sequence of sensor data automatically or semi-automatically. framework, we prototype a decision support system that al- In the car manufacturing process, examples of sensor data lows process engineers to (1) hierarchically compose dynamic include (a) line speeds of the uncoiling machines; (b) CO2 production processes via a GUI and (2) perform deterministic emissions, water consumption, energy consumption, and tem- and stochastic optimization of dynamic production processes. perature of the individual machines or the entire plant; and (c) Users can pose optimization queries against atomic or com- levels of the work-in-progress inventories. This collected data posite process performance models without the need of math- may be filtered and aggregated over time and manufacturing ematical or optimization modeling. The deterministic optimi- levels. In addition, some preprocessing or transformation of zation is implemented by automatic translation of perfor- certainsensordatamaybeperformedtoimprove mance process models into formal optimization models visualization. expressed in Optimization Modeling Language (OPL) and Diagnostic capabilities are needed to detect undesirable solved using the International Business Machines (IBM) deviations from what is considered normal behavior. Int J Adv Manuf Technol Fig. 1 Tesla car manufacturing Detecting such deviations requires continuous testing for any energy consumed and fixing the machine’s parameters, the significant statistical difference between the predicted and ob-