Decision support for sustainable manufacturing using decision guidance query language

Abstract

Sustainability has become a very significant research topic as it impacts many different manufacturing industries. Therefore, the technologies for monitoring, analysing, evaluating and optimising the sustainability performance of manufacturing processes and systems are very critical for decision makers on the shop floor. This paper introduces a decision guidance management system that provides actionable recommendations through quantitative analysis of the sustainability measures of manufacturing processes and systems based on life cycle assessment. The system determines decision preferences through dynamically collected data and decision makers' responses, taking into account the prevailing constraints. Optimal decisions can be derived using mathematical and constraint programming. By using decision guidance query language, this methodology allows users to make optimal decisions without an extensive mathematical or operations research background. Knowledge of relational databases is sufficient for a user to formulate the optimisation problem and obtain optimal solutions. The methodology is demonstrated with a machining operation case study, in which a list of sustainability metrics are identified and sustainability modelling methods are proposed. Important sustainable machining performance measures are optimised, resulting in actionable recommendations.

Keywords:

• sustainable manufacturing
• sustainable machining
• decision guidance management system
• decision guidance query language
• life cycle assessment
• optimisation

1. Introduction

Sustainable development has become an important approach to integrate economic, environmental and social considerations. The US Department of Commerce (DOC) has recently identified sustainable manufacturing as one of its high-priority performance goals, defining sustainable manufacturing as the ‘creation of manufactured products that use processes that minimize negative environmental impacts, conserve energy and natural resources, are safe for employees, communities, and consumers and are economically sound (DOC 2010)’. As such, sustainability-related issues such as energy consumption, emissions and other environmental impact issues are becoming a more integrated part of operational and long-term planning decisions in manufacturing (Kuhl and Zhou 2009).

Machining is one of the major manufacturing operations, involving a number of sustainability indicators that can identify areas with significant potential for environmental impact. These indicators include tool life, coolant and lubricant usage, solid waste and energy consumption. Therefore, the analysis of machining systems and the optimisation of these systems' inputs (control and constraints) and outputs (objective functions) have significant implication for sustainable manufacturing, i.e. the optimisation reduces input sensitivity and provides values for input variables that maximise or minimise an output, given the systems' constraints. It has been observed that the relationship between machining technologies and environmental impact remains insufficiently investigated and the environmental impact due to this manufacturing activity is little evaluated (Narita and Fujimoto 2009). Currently, there are no effective systematic methods for analysing the environmental impact of machining technologies to predict the outcome of its operations and optimising its processes. This paper introduces a decision guidance management system (DGMS) to minimise the environmental impact of manufacturing processes with a case study of machining operations.

Brodsky and Wang (2008) proposed the DGMS data model, which is an extension of the relational database model that can define attributes as random variables with probability distributions, i.e. database tables may have stochastic attributes. A relational tuple assigns numeric values to regular attributes and probability distribution functions (pdf) to the stochastic attributes. The function that defines a pdf for the stochastic attribute is defined using the notion of a dependency graph and transformers. Syntactically, a dependency graph captures the relationship among attributes, including statistical correlation and independence. A transformer is a programme that computes outputs from inputs, relating the values of these attributes. A unified decision guidance query language (DGQL) is an extension of structured query language (SQL). DGQL supports: (1) the integration of querying the data collection and construction of learning sets; (2) learning from the learning sets, using parameterised transformers and optionally defining an estimation utility, such as minimising the sum of squares of errors; (3) probabilistic prediction and simulation using expressions that involve random variables, such as expectation, variance and probability of a logical formula and (4) formulating and solving stochastic or deterministic optimisation, where search space is defined as a set of feasible non-deterministic query evaluations.

The domain knowledge for all these tasks is expressed in DGQL, so that the development of models is as simple as the development of a database reporting application. The DGQL generates the corresponding mathematical models, such as mixed integer linear programming (MILP) or constraint programming (CP) models at run-time, and applies a variety of meta-optimisation heuristics and commercial optimisation solvers. Thus, it combines the simplicity and extensibility of database application modelling with the benefits of using optimisation algorithms based on mathematical programming (MP) and CP. These functions do not need to be manually formulated when analysing sustainable manufacturing applications. Rather, they are automatically derived from abstract model views by the DGMS compiler to describe each application components and factors such as machining operations, equipment, tools, products, energy efficiency, emissions control, waste control, sensing and communication, local power generation and storage and renewable resources. Using DGQL, these abstract models are written much like one implements a database reporting application using the query language SQL, which is simple and intuitive for database application developers or business personnel with database skills.

The DGQL is the core technology that can be applied to decision optimisation for sustainable manufacturing. The contributions of this paper are as follows:

1.

Development of a generic methodology of decision guidance for sustainable manufacturing that • identifies sustainability indicators for a manufacturing process, • models the problem using DGQL, • calibrates unknown parameters of the mathematical model, • supports the definition of algorithms for modelling the indicators and • derives optimal solutions to support decision making.

2.

Develop and implement a case study to demonstrate the proposed methodology, which includes models and algorithms for systematic assessment of decision making for a sustainable machining operation. The following activities are included: • development of a dependency model for emissions of machining system operations, • modelling CO2 emissions of all components of a machining system using DGQL and • development of a predictive scenario exploration demonstration.

The rest of the paper is organised as follows: Section 2 describes the sustainability decision problems currently faced by industry. Section 3 discusses the related work. Section 4 introduces the proposed methodology. Section 5 defines the case study. Section 6 provides a conclusion and discussion.

2. Needs for decision guidance

Recently, an extensive study by MIT Solan Management Review found that ‘there is a strong consensus that sustainability is having – and will continue to have – a material impact on how companies think and act’ (MIT 2010). The US manufacturers are being simultaneously pushed and pulled towards sustainability by increasing regulatory restrictions and consumer preference for environmentally friendly products. These regulations also require that such products be manufactured using processes that have minimal negative impact on the environment. Manufacturing must become more environmentally friendly by

•	using less energy and more renewable sources,
•	using less material,
•	using less water,
•	producing less waste,
•	reducing greenhouse gas emissions,
•	using fewer scarce or toxic materials and
•	using less space.

To achieve the above objectives, industry must develop and adopt new technologies and practices. Machining operations are among the more widely used manufacturing processes because many consumer goods are made of machined components. Therefore, increasing the sustainability of these operations can significantly reduce the environmental impact of consumer goods. Decision optimisation modelling for sustainable manufacturing will help industry achieve needed analyses, evaluations and recommendations. However, before accomplishing this, there are challenges that need to be addressed, such as the lack of

•	standardised sustainability indicators and metrics,
•	sources of sustainability information,
•	sustainability measurement and data collection methods,
•	algorithms for modelling sustainability metrics,
•	sustainability analysis and evaluation methodology and
•	modelling methods for decision guidance for sustainable manufacturing.

An incomplete understanding of sustainability in manufacturing processes and inadequate sustainability information prevent decision makers from making major improvement decisions. For example, analysts do not have a complete understanding of the relationship between machining technologies and environmental impact, as well as the interaction among the different inputs to the machining operations. This situation prevents decision making that would improve the sustainability of the operations. Scientists and researchers have carried out analysis of the environmental impacts due to manufacturing processes and provided technologies to improve the sustainability performance for individual situations. However, an ad hoc and piecemeal problem-solving approach does not guarantee a continuous and complete solution for sustainable manufacturing. The lack of decision guidance makes it a particular challenge for human decision makers. Therefore, data analysis and decision support are important for choosing the best (most sustainable) course of action in material, technology, equipment and manufacturing process selection and design. Systematic methodologies and tools that provide quantitative analysis and actionable recommendations are needed for analysing and evaluating the sustainability performance and aiding the decision-making process.

3. Related work

Sustainable machining has recently attracted considerable attention from researchers. However, the accomplished research reported on this topic is still insufficient. Approaches for solving the problem have varied from finding alternative material removal methods or changing of material to focusing on optimising the various inputs to the machining without changes to technology. The objective is to optimise environmental factors such as use of material, cutting fluids and energy. In the research conducted on environmental analysis of machining operations from a systems' perspective (Dahmus and Gutowski 2004), the machining system operations include activities such as tool preparation, material production, material removal and cleaning, among others. Akbari et al. (2001) carried out the life cycle of machine tools from manufacture and use to disposal. During machining, it was found that more energy may actually be used for running peripheral devices such as coolant pumps, hydraulic pumps and control devices than for actual cutting.

A system for evaluating the environmental burden of a machine tool operation has been developed (Narita et al. 2006). It extends previous methods that evaluate the difference among dry, minimal quantity lubricant (MQL) and wet machining approaches and includes factors such as depth of cut, feed rate, spindle speed and tool path pattern to optimise the operations for least environmental burden. The environmental burden is evaluated after converting all measures to CO₂ equivalents. Narita and Fujimoto (2009) developed algorithms to calculate the environmental burden due to machine tool operations. They calculated the environmental impact of dry, MQL and wet machining using the five impact categories of global warming, acidification, eutrophication, photochemical oxidants, human toxicity and ecotoxicity. However, a methodology that provides a systematic approach to aid decision making for optimising environmental impact of machining operations was not addressed.

Previous research has been conducted in the use of simulation modelling to optimise sustainable manufacturing decisions. It includes Heilala et al. (2008) who describe an interactive tool that jointly combines simulation models and data to make human- and environmentally friendly decisions for sustainable manufacturing systems. But the final decision depends on human judgement. The tool was developed for investment and manufacturing decision making for products in the early design stage. Solding et al. (2009) describe an integrated system that combines simulation and an energy analysis tool, Method for analysis of INDustrial energy systems (Tari and Söderström 2002), to optimise energy use in foundry plants. Four case studies are used to demonstrate the approach. Johansson et al. (2009) present a simulation model of an automotive paint shop that simulates different input parameter options to determine the one with least CO₂ emission. Jayal et al. (2010) overview the trends, new concepts and analysis approaches used in the development of sustainable products, processes and systems. This overview shows that previous research includes systems perspective considerations for machining operations and the analysis of sustainability in terms of safety/health, work environment, material and power usage and waste management. Material removal technologies are also reviewed. The modelling of machining operations shows the use of finite element methods and the analysis of chip formation mechanisms for various materials. The overview shows that the modelling of various inputs and settings such as work piece geometry and size, cutting speed and tool path to minimise the environmental impact of the operations is lacking in the literature. Therefore, none of previous research approaches focuses on a guidance system that can provide recommendations for human decision making.

Brodsky and Nash (2005) proposed and implemented the language CoJava, which offers both the advantages of simulation-like process modelling and the capabilities of true decision optimisation. The syntax of CoJava extends Java with special constructs to (1) make a non-deterministic choice of a numeric value; (2) assert a constraint and (3) designate a programme variable as the objective to be optimised. A CoJava programme thus defines a set of non-deterministic execution paths, each being a programme run with a specific selection of values in the choice statements. The semantics of CoJava interprets a programme as an optimal non-deterministic execution path, namely a path that (1) satisfies the range conditions in the choice statements; (2) satisfies the assert-constraint statements and (3) produces the optimal value in a designated programme variable, among all execution paths that satisfy (1) and (2). Thus, to run a CoJava programme, first find an optimal execution path, and then procedurally execute it. A reduction was developed to a standard constraint optimisation formulation. Constraint variables represent values in programme variables that can be created at any state of a non-deterministic execution. Based on the reduction, a CoJava constraint compiler was also developed. The compiler operates first by translating the Java programme into a similar Java programme in which the primitive numeric operators and data types are replaced by symbolic constraint operators and data types. This intermediate Java programme functions as a constraint generator. This programme is compiled and executed to produce a symbolic decision problem. The decision problem is then submitted to an external optimisation solver. Brodsky and Wang (2008) developed the DGMS methodology, in which they use a similar concept except that they replaced Java programming with a relational database modelling, which is much easier for many business-oriented users. In this paper, we apply the DGMS methodology to sustainable manufacturing problems.

Desmira et al. (2009) performed environmental burden analysis of high-speed milling machining. They used the environmental burden analyser to calculate the total CO₂ emissions for different cutting speeds, and then using experimental data and the mathematical least squares method, they derived a mathematical function, in which the total CO₂ emissions was the dependent variable and spindle speed was the independent variable. They determined the optimised spindle speed for a minimised total CO₂ emission. In order to accomplish this, researchers or analysers need to have strong mathematical knowledge or operations research (OR) background. The same case scenario and test data are used for DGQL modelling in this paper. DGQL can help information technology (IT) or business users to formulate the optimisation problem and submit it to an external mathematical solver, which then automatically returns an optimal solution.

4. Proposed methodology

Discrete event simulation (DES) has been used as a tool to model manufacturing processes for ‘what-if’ analyses of various scenarios to aid decision making. However, the output of simulation does not provide optimal solutions or any specific recommendations. Making decisions based on the simulation results for sustainable manufacturing could still be a challenge for a human user, requiring considerable time to make multiple runs and analyse the output data to obtain useful information. Therefore, a decision guidance system is needed for easy, faster and more reliable decision making. A methodology that provides actionable recommendations for sustainable manufacturing is proposed as illustrated in Figure 1. It involves identifying the manufacturing process of interest, in which decisions need to be made, collecting domain knowledge for the relevant indicators and their metrics, developing data models of the manufacturing process, e.g. dependency graph and database schemas, collecting and processing available required data, calibrating the model for unknown parameters, and optimising the decision indicators. The methodology's steps are addressed in detail in the following sections.

Figure 1 Decision guidance process for sustainable manufacturing.

4.1 Problem formulation

To form a problem under study, the following steps should be used:

•	Identify the manufacturing process to be optimised, e.g. a machining process.
•	Define the objective(s) of the study, e.g. to find optimal cutting conditions that lead to minimal total CO2 emissions of the process.
•	Define the scope of the study including scenarios, assumptions and level of abstractions.
•	Determine at a high level what types and forms of information are needed by the decision makers, e.g. the cutting speed range to achieve minimal environment impact. This, in turn, aids decision-making regarding new investments in a machine tool.
•	Determine what the decision factors of the problem are, e.g. cutting speed of a machine, costs of a machine.

4.2 Knowledge collection

In addition to the traditional manufacturing process information, sustainable manufacturing-specific information on environmental, economic and social aspects is also needed. In this paper, we focus mainly on the environmental impact. As such, we need to

•	Thoroughly understand the problem under study, e.g. the machining operation, machining conditions, input requirements to produce the part, finished product and by-products and most importantly, the environmental impact of the machining process.
•	Identify key performance indicators, which may not be sustainability indicators e.g. cycle time, throughput.
•	Identify the sustainability indicators and their metrics for the manufacturing processes.
•	Identify the relationships among the indicators.
•	Identify data sources for sustainability information, e.g. actual machine, machine specifications, life cycle inventory (LCI) database (EPA 2010) and other manufacturing applications.
•	Identify or develop algorithms to model the problem.
•	Identify data needed for the decision guidance modelling, and based on the variables in the algorithms, determine which data are available and which are not.

Determining the appropriate metrics of sustainability indicators and computing them for sustainable manufacturing are necessary steps to enable companies to quantitatively measure the sustainability performance in a specific manufacturing process. According to Greiner (2001), there are numerous lists of environmental performance indicators. But these lists provide little insight into how manufacturers might revise these indicators to better measure sustainability. The Lowell Center for Sustainable Production at the University of Massachusetts has developed a tool to enable companies to evaluate the effectiveness of sustainability indicator systems. The tool consists of five levels for categorising existing indicators related to the basic principles of sustainability, i.e. (1) facility compliance/conformance indicators; (2) facility material use and performance indicators; (3) facility effects indicators; (4) supply chain and product life cycle indicators and (5) sustainable systems indicators (Greiner 2001). Wichita State University maintains an LCI database that provides data to assess a product life cycle at the manufacturing stage; the focus is on manufacturing unit processes (UPLCI 2010). A life cycle heuristic is to establish representative estimates of energy and mass loss from a unit process in the context of manufacturing operations for products. The unit process life cycle inventory (UPLCI) profile is defined as the use of processes that generally have high automation and are at the medium-to-high-throughput production compared with all other machines/equipment that perform a similar operation. This is consistent with the life cycle goal of estimating energy use and mass losses representative of efficient product manufacturing. The National Institute of Standards and Technology (NIST) efforts (Feng et al. 2010) on the development of standard indicators and metrics sets for products and processes sustainability performance can also help in this step.

4.3 Data model

The data model provides a mechanism for describing sustainability data in the scope of the study. Based on the data model, the answers for the following typical questions can be directly or indirectly obtained:

•	What sustainability indicators are we interested in modelling?
•	What are the metrics of each indicator and what measurement method is used?
•	What are the data requirements in terms of contents, units and formats?
•	What are the dependency relationships (e.g. correlations) among the indicators and their attributes in the database?
•	Will the data come from a real-world or virtual world (e.g. simulation) data source?
•	How should the data be collected?
•	How should the data be processed?
•	How can the data be expressed in the data model?
•	What are the exact relationships in mathematical form?
•	What are the relational constraints?
•	What are the data types of the attributes in the database table?
•	What are the attributes of each indicator in the database table?
•	What are the constraints imposed on the attributes?

The dependencies among attributes are important for stochastic relational models. The dependencies can be modelled using a dependency graph, which is composed of a set of connectors that have input and output attributes representing random variables. A connector in the dependency graph may have an associated transformer, in which numeric variables are polymorphic and correspond to either a numeric value or a random variable with a particular pdf. Also, some parameters in a transformer may be declared as learned, to indicate that they are unknown at the time of transformer's definition, but can potentially be learned if a learning set is provided. Thus, a connector and a corresponding transformer represent partial knowledge held by the user on how output random variables are computed from the inputs. This will be further explained below in the case study section using an example.

4.4 Data acquisition

Data are crucial for performing successful analysis and providing decision guidance, Skoog's analysis of DES showed that about 31% of the total modelling time is used for gathering, extracting and processing data, and he introduced methodologies for input data management (Skoog 2009). Currently, in most cases, sustainability data are not being identified and collected. Even when these data are collected, they are often not stored in a relational database that could be used directly by DGMS. Sometimes, the data are indirectly stored and need to be computed. For example, some measures are a total amount derived from a calculation and some measures need to be converted to other measures. Normally, the analysis of sustainability and life cycle assessment (LCA) requires LCI data and characterisation data as inputs (EPA 2010). These data need to be identified, collected and processed. Heilala et al. (2008) listed the general information needed for sustainable manufacturing as

•	energy consumption, direct and indirect,
•	CO2 emissions, direct and indirect,
•	other air emissions, e.g. NO x and VOC,
•	solid waste including hazardous waste,
•	water emissions and
•	toxic material used in every process step, e.g. cutting fluids.

The sources of relevant sustainability data need to be identified; it can be static data from LCI databases or dynamic data from either the real or the virtual world. The following sections describe the different sources and collecting methods of data for the purpose of decision optimisation (Shao et al. 2010b).

4.4.1 LCA tools and LCI data

LCA is a powerful tool for analysing the sustainability of manufacturing systems, using multiple commensurable (i.e. measurable in similar units) aspects of quantifiable systems, based on systems thinking. The LCA parameters include energy, resources, waste and emissions. Results from using an LCA tool provide guidance about the relative impacts of different types of products, materials, services or industries with respect to resource use and emissions throughout the supply chain. For example, the environmental effect of producing an automobile would include not only the impacts at the final assembly facility but also from mining metal ores, making electronic parts and forming windows, etc. that are needed for parts to build the car (EIO-LCA 2010). The LCI data are needed for LCA, there are two data streams: (1) primary (foreground) data and (2) secondary (background) data. Primary (foreground) data are derived directly from the process in question; these are the most accurate data that can be applied to an LCA, e.g. bill of materials, logistics and usage patterns. In many cases, foreground data are proprietary and are not readily available to the public and it becomes necessary to use background data such as databases and data from peer-reviewed literature, etc. High-level assessment can depend on general databases such as Ecoinvent, GaBi, and SimaPro (Kent 2007, Stutz 2010). Currently, there are publicly available LCI databases from different regions. An example of such a database is the European Commission life-cycle database (ELCD). The ELCD core database comprises life cycle emission and resource consumption data from front-running EU-level business associations and other sources for key materials, energy carriers, transport and waste management. The focus is on data quality, consistency and applicability. The respective data-sets are officially provided and approved by the named business association (European Commission 2010).

4.4.2 Real-world data

Real-world data are the data that can be collected from devices, machines, manufacturing processes, databases, or other application systems. For example, the sustainability data from real world may be extracted through the following systems:

•	MTConnect (Vijayaraghavan et al. 2008) is a middleware standard that provides the capability to extract data from machine tools using the eXtensible Markup Language standard. MTConnect provides the mechanism for system monitoring, process and optimisation with respect to energy and resources. The information on efficient use of machines and systems in the production facility (including use of consumables such as water and fluids) will be valuable to designers of tools, processes, and systems to develop sustainable manufacturing systems and facilities. Bengtsson et al. (2010) developed a case study of using MTConnect to acquire Boeing production data for sustainable machining modelling.
•	The object linking and embedding for process control (OPC) is an existing technique for monitoring manufacturing systems and their status (OPC 2009). The OPC standards specify the communication of industrial process data, alarms and events, historical data and batch process data between sensors, instruments, controllers, software systems and notification devices.
•	Energy management system (EMS 2010) enables energy consumption data monitoring and collection. The energy data can then be mapped to the logged data from other sources such as programmable logic controller (PLC) and OPC. It can be, for example used to identify potential energy peaks for certain operations within manufacturing.

4.4.3 Virtual world data

Virtual world data for sustainability can be obtained from a simulation system, design system, engineering system or product data management system. For example,

•	CAD/CAM systems. SolidWorks Sustainability, for example provides LCA on parts or assemblies within the design window and compares materials for their environmental impact.
•	Simulation systems. Environmental impact data can be obtained from the output of a simulation system (Shao et al. 2010a). A machining model can be developed to execute numerical control (NC) programmes (Shao et al. 2003). Simulation output data can be used to create a training set for DGMS.

4.5 Decision Guidance Query Language (DGQL)

A typical DG database contains a regular relational database, which is dynamically modified by users and other systems. As introduced in Section 1, DGQL provides stochastic or deterministic transformers to define new attributes, which may be random variables. Given a random variable in one of the attributes using the ‘AUGMENT’ statement in DGQL, which augments a table with additional attribute columns that are populated with non-deterministic decision variables, DGQL allows a random selection of a value according to its pdf. It also provides operations on random variables, including computing their expectations and standard deviations, and the probability of a logical formula. Over these tables, the user can write queries in a conventional manner as if the variables have already been instantiated. Expressions can be used in the SELECT and the FROM clauses of DGQL.

DGQL also allows users to define objective functions over the augmented table using regular query language and uses the ‘INSTANTIATE’ statement to instantiate the variables inserted by the ‘AUGMENT’ statements with the objective value optimised. An optimisation DGQL query also specifies an aggregation function, such as the sum on a number of attributes, defined in the SELECT clause, to be minimised or maximised. The optimisation semantics of DGQL are to (1) find an optimal non-deterministic query evaluation, i.e. one that produces the minimal or maximal (as requested) aggregation answer, and then (2) compute the query with values for the non-deterministic attributes corresponding to the optimal evaluation path. The search space in the DGQL optimisation queries is defined implicitly through a set of non-deterministic query evaluations, rather than by arithmetic constraints over numeric constraint variables. The transformer is used in the optimisation query in the computation of the aggregate, and its learned parameters are interpreted as non-deterministic values to be optimally instantiated. MP techniques are used behind the scenes to instantiate the variables (Brodsky et al. 2009).

The MP tool is a concept coming from the world of OR. A typically MP solver is IBM ILOG/CPLEX, to which problem instances say represented in A Mathematical Programming Language (AMPL) will be submitted. ILOG CPLEX is an optimisation software package. ILOG CPLEX solves integer programming problems, very large linear programming problems, quadratic programming problems and problems with convex quadratic constraints (solved via second-order cone programming or SOCP). It has a modelling layer called Concert and is available with several modelling systems like AMPL (CPLEX 2010).

To develop any DGQL model for sustainable manufacturing, the following questions need to be answered:

•	Based on the objective of the model, what should the constraints of the model be?
•	What are the available data?
•	What are the deterministic data?
•	What are the stochastic data?
•	What are the data needed but not available?

Some of the knowledge for deterministic or stochastic models, such as spindle speed vs. cutting time model, may not be known a priori. However, there may be historical, experimental and statistical data that can be leveraged to ‘learn’ the unknown parameters, using regression analysis (RA). Mathematical abstractions can be used to describe the learning problems, whereas the same domain knowledge is being used for both optimisation and learning. To support learning, a DGQL query defines a transformer programme with parameters designated as ‘learned’. The transformer is associated with an optimisation DGQL query that computes a learning set and from it, an aggregate, such as the sum of squares of errors, is minimised. RA is one of the metamodelling techniques for investigating and modelling the relationship between variables. As an input to RA, a parametric functional form can be linear, e.g. , or nonlinear and a training data-set is needed, e.g. tuples of the form , where f is an experimental observation of the function f-value for an input set of . The goal of RA is to find the unknown parameters, e.g. that ‘best approximate’ the training set. Once we find the optimal set of the parameters, we replace the variables in the template with the optimal parameter values (Shao et al. 2009).

4.6 Actionable recommendations

Decision guidance is needed to make decisions that improve performance, lower cost and make manufacturing operations both economically and environmentally sustainable. Examples of such decisions include finding machine cutting conditions that minimise CO₂ emissions, taxation policies for shop floor pollution, and making public policies guided by the most positive outcomes. In these applications, decision guidance is usually provided in the presence of large amounts of dynamically collected data. The technical tools needed for decision guidance typically include MP to optimise decisions. Such tools require considerable programming skills and efforts. On the other hand, using DGQL, users do not need to have mathematical and OR background. Only relational database knowledge and domain knowledge are needed to formulate the optimisation problem and obtain the optimal solutions. Essentially, each such model is comprised of table schemas that hold the relevant information and the DGQL views that compute the relevant sustainability metrics, such as energy consumption, carbon emissions and cost. Thus, these models can be used for asking what-if analysis queries. DGQL views can be annotated to indicate that the value it computes is to be used as an optimisation objective. The DGQL optimisation engine then parses all of the relational database view definitions and builds an algebraic representation in memory. This includes issuing multiple queries against the database to retrieve all the model parameters from the various tables. Once this algebraic model is built, DGMS automatically generates, at run-time, a formal MP problem with mathematical equations, inequalities and the objective function and deploys a mix of algorithms best suited for the problem at hand, e.g. MILP using ILOG CPLEX solver. Optimal values of the decision variables will be derived and returned to the database table. These optimal values provide decision makers with actionable recommendations for the problems. Therefore, when a new factor for sustainable manufacturing applications is introduced, the only requirement is to add a simple SQL-view like model – DGQL, whereas all the learning, optimisation and other DGMS functions are automatically implemented with the use of the DGMS compiler. All systems' static and dynamic data will be normalised and stored in a relational database management system (RDBMS). Thus, the goal is to create a flexible, yet normalised design of database schemas that can be maintained by the companies' database administrator.

5. Case study

To validate the proposed methodology, a case study on decision guidance for sustainable machining in a shop floor environment has been defined and implemented. This case study activity includes: (1) identifying the sustainability indicators for a machining operation; (2) identifying the performance metrics for the identified indicators, e.g. the algorithms for calculating the CO₂ emissions; (3) developing a dependency graph model of the identified sustainability indicators; (4) defining optimisation objective, e.g. minimum total CO₂ emissions; (5) developing the DGQL model for the problem and (6) finding the optimal solutions, e.g. optimal spindle speed and cutting conditions.

5.1 Case scenario

The scenario uses a machining centre with a 10 mm diameter carbide square coating cutting tool. The flute number is six and the number of teeth is two. The axial depth is 0.1 mm, the radial depth is 10 mm and cutting length is 60 m. The objective is to determine the appropriate cutting conditions, i.e. the optimal machining speed to minimise the total CO₂ emissions for the machining system operation. This case scenario was described in Desmira et al. (2009), DGQL is used to perform the decision optimisation. To assess a manufacturing process efficiently in terms of environmental impact, the concept of a unit process is applied. The unit process consists of the inputs, process and outputs of an operation. Each unit process converts material/chemical inputs into a transformed material/chemical output. The transformation of input to output generates five UPLCI characteristics (Kalla et al. 2009):

•	input materials,
•	energy required,
•	losses of materials (that may be subsequently recycled or declared waste),
•	major machine and material variables relating inputs to outputs and
•	resulting characteristics of the output that are often input to the next unit process.

Figure 2 illustrates an overview of the case study, which shows inputs and outputs for machining modelling.

Figure 2 Case study overview diagram.

To model the problem, we need to mathematically capture the emission as a (possibly stochastic) function of the decision choices (selection of machine tool, cutting tool, product, cutting speed, etc.), i.e. we need expressions for the emissions model due to electricity consumption (as a stochastic function of machining time, etc.), coolant/lubricant usage (as a stochastic function of machining time, etc.), cutting tool (as a function of tool wear and cutting speed) and metal chips. We use algorithms developed by Narita and Fujimoto (2009) and list the algorithm that calculates the equivalent CO₂ emission of the sustainability impacts for metal chips as an example below.

Metal chip generates a sustainability impact that can be calculated based on chip weight as follows:

where CHe is the sustainability impact of metal chips in carbon dioxide equivalents (kg-CO₂); WPV the stock volume (cm³); PV the product volume (cm³); MD the stock material density (kg/cm³) and WDe the sustainability impact of metal chip processing (kg-CO₂/kg).

The sustainability data are collected using MTconnect, from a machine specification, LCI database, simulation system output and other databases.

The LCA functional unit and system boundaries are the key elements of LCA. The functional unit is a measure of the function of the studied system and it provides a reference to which the inputs and outputs can be related. This enables comparison of two essential different systems (ISO 2006). The functional unit for the machining system in this case study is defined as ‘for one specific part, for a given NC programme, with given material of the part and the weight of the stock’. The system boundaries determine the unit processes to be included in the LCA study. We consider the potential environmental impacts throughout a product's life-cycle (i.e. cradle to grave) from raw material acquisition through production, use and disposal. The products are machined part, cutting tool, coolant and lubricant oil. LCI data are based on the data in Narita and Fujimoto (2009).

All the processed data can be structured based on the data model (dependency graph). For one specific part, on a specific machine, with given material of the part and the weight of the stock, some of the questions a decision maker may ask include

•	What is the ideal spindle speed for having minimum total emissions?
•	What is the machining time based on the spindle speed?
•	What are the cutting conditions (cutting speed and feed rate) for the selected spindle speed?

DGQL modelling formulates the optimisation problem and submits the problem to an external solver, such as ILOG CPLEX, to obtain the optimal solutions. Conclusions can be drawn and suggestions can be made to the factory floor for better decision making based on the results. A simulation model of the same process can serve as a validation mechanism.

5.2 Dependency graph and transformers

Based on the algorithm discussed in Section 5.1, we developed a dependency graph that captures the relationship among attributes, including statistical dependencies. The dependency graph also guides the design of transformers that relate the values of these attributes. Figure 3 shows a simplified version of the dependency graph for the total emissions. Circles correspond to attributes, such as cuttingSpeed, toolLife and totalEmission. Boxes correspond to dependence connectors, such as Emission and Waste. Single-line boxes correspond to deterministic connectors. For example, the waste connector suggests that the value of its output attribute matelChipEmission is deterministically dependent on its input attributes stockVolume, productVolume, materialDensity and the sustainability impact from production and disposal. Double-line boxes correspond to stochastic connectors. For example, dependence connector CSpeed, as shown in Figure 3, suggests that its output attribute, machiningTime, is stochastically dependent on the values of its input attributes spindleSpeed, which is the optimal input variable in this case.

Figure 3 Dependency graph model for cutting speed vs. CO₂ emissions.

5.3 Relational schemas

Based on the scenario and the dependency graph, we can define the following input database table schemas.

•	Product (pID, pName, pVolume, maID) represents the name (pName), material (maID) of a particular product (pID) and its volume (pVolume).
•	Stock (sID, sVolume, maID) represents a particular material (maID) of a particular product (sID) and its volume (sVolume).
•	Material (maID, maName, mDensity, mWDe) represents the material name (maName), density (mDensity) and sustainability impact (mWDe) for processing a particular material (maID).
•	Machine (mID, mName, tID, cID, cL, cC, cAc, cWAq, cAwaq, lID, lSi, lSv, lLi, lLv) represents the machine name (mName), tool (tID) used on a particular machine (mID), a particular kind of lubricant (lID) and coolant (cID) being used on that machine, mean interval of coolant update (cL), initial coolant quantity (cC), additional supplement quantity coolant (cAC), initial water quantity(cWAq), additional supplement water quantity (cAwaq), mean interval between discharges (lSi) and spindle lubricant discharge rate (lSv), and mean interval between discharges (lLi) and slide way lubricant quantity (lLv) for lubricant (lID).
•	Tool (tID, tD, tFPT, tNT, tM, tW, tL, tRn, tPe, tDe, tRe) represents the tool diameter (tD), feed per tooth (tFPT), the number of teeth (tNT), material (tM), tool weight (tW), tool life (tL), number of regrinding (tRn) for a particular tool (tID), and sustainability impact from production (tPe), from disposal (tDe) and from regrinding (tRe).
•	MachineUnitElectricityConsumption (mID, mSME, mSPE, mSCE, mCME, mCPE, mTCE, mATCE, mMGE, mSBE, eK) represents the machine unit electricity consumptions for servomotor (mSME), spindle motor (mSPE), spindle cooling system (mSCE), compressor (mCME), coolant pump (mCPE), chip conveyor (mTCE), auto tool changer (mATCE), tool magazine motor (mMGE), standby energy (mSBE) for a particular machine (mID) and CO2 emission intensity (eK).
•	Coolant (cID, cType, cPe, cDe, cWAe) represents the type of coolant (cTytpe) for coolant (cID), its sustainability impact from production (cPe), from disposal (cDe) and from water distribution (cWAe).
•	Lubricant (lID, ltype, lSPe, lSDe, lLPe, lLDe) represents the type of lubricant (lTytpe) for lubricant (lID), spindle sustainability impact from production (lSPe), from disposal (lSDe) and slide way sustainability impact from production (lLPe) and from disposal (lLDe).

From the above inputs, we need to compute the following results as output, which minimise the computed CO₂ emissions for the defined machining process on the shop floor:

•	SpindleSpeed (mID, tID, pID, sID, sSpeed) represents the spindle speed (sSpeed) on a particular stock (sID) for the product (pID) being produced using a particular tool (tID) on a particular machine (mID).
•	MachiningTime (mID, tID, pID, sID, mTime) represents the machining time (mTime) on a particular stock (sID) for the product (pID) being produced using a particular tool (tID) on a particular machine (mID).

5.4 DGQL views

Figure 4 shows the screen capture of DGQL views and queries in a Postgres database. Postgres is a powerful, open source object-relational database system. It provides a development platform in which to develop in-house, web or commercial software products that require a capable RDBMS (Postgres 2010).

Figure 4 DGQL queries for the case in the Postgres database.

5.5 Calibrate unknown parameters

As shown in Figure 3, the only control variable is spindle speed. Machining Time and Tool Life are affected by varying spindle speed. The rest of the parameters are constants because they are mainly LCI data and specification data. To simplify the problem, all the emissions except from the cutting tool are combined, i.e., only two components need to be modelled. The first is CO₂ emissions due to the cutting process, i.e. E _p = electricityConsumptionEmission+coolant Emission+lubricant Emission, and the second is the emissions resulting from the cutting tool, i.e. E _c = cutting Tool Emission. These two components make up the totalEmission, i.e. E _T = E _p+E _c. We know that, in general, the higher the cutting speed, the shorter is the machining time needed, and, in turn, the lower the amount of the CO₂ emitted E _p during the machining process. However, on the other hand, as the speed increases up to a point, the emission E _c from the cutting tool goes up due to rapid tool wear. Thus, overall CO₂ emission gradually increases after a point. Therefore, the unknown parameters such as cutting tool wear and machining time have to be learned using RA. We used data in Figure 5 as learning data, it is an empirical model based on the data derived from Desmira et al. (2009). RA is used to identify the coefficients of a quadratic approximation of the graph. Once we find the optimal parameters, we can replace the variables in the formula with the values.

Figure 5 Experimental emission data from both process and cutting tool.

5.6 Optimal solution

This case study is a simple deterministic optimisation problem, our goal is to determine the input parameter (spindle speed) that corresponds to the minimum CO₂ emissions. In Sections 5.3 and 5.4, we showed the database encoding of CO₂ emissions due to the machine process. The implementation is built on the relational database PostgreSQL that allows us to use standard database techniques to model the problem. In fact, we can validate and verify the model by manually setting the values in the SpindleSpeed table and see if the empirical value corresponds to that obtained by querying the Total_Emission computation view.

To turn this database programme into an optimisation problem, the DGQL statement for minimising the total emission has to be applied as follows:

SELECT * FROM minimise (Total_Emission)

Once the solver has obtained a solution, the optimal input values are retrieved and inserted back into the database. The optimal results can provide actionable recommendations, e.g. the ideal speed for lowest emissions and the optimal tool utilisation. In this study, the optimal spindle speed is 6400 rpm. DGQL also calculates the cutting conditions, i.e. the cutting speed and feed velocity. The computations are based on the following formulae:

where V _c is in metres per minute, the diameter d is in millimetres and n is in revolutions per minute.

where V _f is in millimetres per minute, f _z is the feed per tooth and z is the number of teeth.

The conclusion is that the cutting conditions for the lowest possible total emission are with the spindle speed of 6400 rpm, cutting speed of 200.96 m/min, and feed rate of 1280 mm/min. This leads to a recommendation of an investment in a high-speed machine tool. Figure 6 shows the DGQL views of optimisation and calculations of the cutting conditions.

Figure 6 DGQL views (optimisation and calculations of cutting condition).

6. Conclusion

This paper has introduced a generic systematic methodology to help decision making for sustainable manufacturing and applied the methodology to a machining operation case study to determine optimal cutting conditions that minimise the total CO₂ emissions from the machining system. The procedures for performing the DGMS activity using the methodology include the identification of relevant sustainability indicators and their metrics, the data sources, the methods for collecting and processing the identified information, and the development of a data model. By using DGQL modelling, the optimisation problem can easily be formulated and solved by an exterior solver. The optimal model outputs provide actionable recommendations to decision makers for the improvement in sustainability on the shop floor. The case study has shown that the machining centre purchase decision can be made based on the predicted cutting speed that lead to the minimal total CO₂ emissions. The model recommends a high-speed machining centre. Compared to traditional optimisation methods, DGQL makes decision optimisation for sustainable machining more intuitive and user friendly for IT and business personnel.

In the future, a library of reusable DGQL models, algorithms and tools will be developed for assessment and decision making on both operational- and investment-planned aspects of sustainable manufacturing. The models will be used as a template with minimum adjustment for other similar cases. New types of applications can be dynamically added to the template library. The existing types of applications within the library can be executed with new data, so that different companies that have the same problems could use the template by inputting their specific data in order to obtain optimal decision outputs. The case study discussed in Section 5 will be modelled using a different approach, for example DGQL can solve more complex computations with this emissions model, e.g. with a nonlinear model of lubrication emissions based on fluid dynamics by having additional database tables and views defined. More case studies for sustainable manufacturing will be implemented. The DGMS may also be integrated with DES. DGMS works using a top-down approach, whereas DES works using a bottom-up approach. DES can model some parts of a manufacturing system and provide DGMS with a learning data set. There are also opportunities for developing a common conceptual information model that can be used by both DES and DGMS, so that both methodologies share the analysis and the data. DES results of different scenarios can be used by DGMS to determine an optimal solution based on the goals of the decision maker. DGMS can also provide a validation mechanism for DES models.

Disclaimer

No approval or endorsement of any commercial product by the NIST is intended or implied. Certain commercial software systems are identified in this paper to facilitate understanding. Such identification does not imply that these software systems are necessarily the best available for the purpose.