ENABLING THE DIGITAL FACTORY THROUGH THEINTEGRATION OF DATA-DRIVEN AND SIMULATION MODELS

Olivia J. Pinon*, Dennis JL Siedlak*, Dimitri Mavris**Aerospace Systems Design Laboratory (ASDL)

School of Aerospace EngineeringGeorgia Institute of Technology, Atlanta, GA, 30332-0150,

USAKeywords: Digital factory, Digital twin, Data-driven modeling, Disruption simulation, Production

scheduling

Abstract

The digital factory relies on the development ofadvanced integrated analysis and simulation ca-pabilities that better utilize the vast amounts ofdata produced by production systems. This paperdiscusses the integration of data-driven and sim-ulation models in the context of a partially auto-mated aircraft assembly process, with the goal toanalyze and assess the impact of disruptions onthe production system and identify strategies tobest recover from these disruptions.

1 Motivation

Manufacturing is a very established industry andyet factories are continuously evolving and be-coming more complex and dynamic. Today, largenumbers of sensors gather process data (time,energy consumption, environmental conditions,etc.) throughout the factory. This digital revo-lution, often defined as Industry 4.0 and drivenby advances in big data analytics, artificial intel-ligence, robotics, augmented reality and additivemanufacturing, will change the nature of design,manufacturing and the interactions between de-signers, manufacturers, suppliers, customers, andphysical industrial assets [3, 13].

1.1 The Digital Factory

While the concept of a digital factory is not new(it has been introduced in the literature as early as

1950), it is now, more than ever, within reach ofbeing achieved. Enabling the concept of the dig-ital factory, however, is not without challenges.In particular, it requires the development of ad-vanced integrated analysis and simulation capa-bilities that better utilize the vast amounts of dataproduced. As discussed by William P. King,chief technology officer of the Digital Manufac-turing and Design Innovation Institute (DMDII),“manufacturing generates more data than anyother sector of the economy” [15]. In a digi-tal world, data is a strategic asset. If properlycaptured and used, data has the potential to un-lock new sources of revenues [3] and enable sig-nificant improvements in innovation, product de-sign, operational effectiveness, reliability, time-to-market, customer satisfaction, and sustainabil-ity. While the benefits and competitive edgebrought by the digital thread/digital twin, and bigdata technologies in particular, are widely dis-cussed and fully acknowledged within the com-munity [23, 8], achieving these benefits is notwithout challenges.

As highlighted by Baur and Wee [3], “manymanufacturing companies have deep expertise intheir products and processes, but lack the exper-tise to generate value from their data.” Hence,while many seamlessly integrated hardware, soft-ware and technology-based services exist that en-able the digital factory concept and the genera-tion of data at any level of detail desired, im-proved frameworks and platforms that organize

1

OLIVIA J. PINON*, DENNIS JL SIEDLAK*, DIMITRI MAVRIS*

data-analytic thinking and facilitate data-drivendecision making across the many facets of thebusiness are still needed [24].

1.2 A Paradigm Shift

To be effective, the digital factory needs to gobeyond tracking and collecting data, and insteadfocus on transforming this data into intelligence[28] to help make plants more transparent andtruly proactive. Doing so requires a shift in theway we approach and model factories. Tradi-tionally, data is collected from the real factoryand analyzed to help identify important processparameters, inefficiencies, etc. A separate ef-fort then consists in building simulation modelsto help predict the state of the factory under dif-ferent scenarios and eventually support decisionmaking. To unleash the real value of data, oneneeds to enable the link between data and thedigital twin (Figure 1), i.e. integrate data-drivenmodels with simulation models. Doing so willenable the information and knowledge containedwithin the data to be integrated within the manu-facturing system [9].

1.2.1 Opportunities for Machine Learning Ap-plications

The rise of automation has enabled huge amountsof data to become available. For all purposes, thisdata, due its size and the many variables it cap-tures, cannot be analyzed using traditional (man-ual) techniques. However, the continuously in-creasing amount of data available, as well as itshigh dimensionality and variety (sensor data, en-vironmental data, machine tool parameters, etc.)[6, 27] lends itself particularly well to the use ofmachine learning algorithms [27, 20] and the de-velopment of data-driven models for prediction,detection, classification, regression, or forecast-ing [5, 12]. As such, machine learning algorithmshave successfully been applied to support predic-tive manufacturing [16, 17], manufacturing pro-cess monitoring and control [7, 2, 1, 25, 11], etc.

1.2.2 Modeling of Production Systems

The integration of data and information withsimulation models is traditionally a challenging,time-consuming task [10]. As discussed byFowler and Rose, the time to “design, collectinformation/data, build, execute, and analyzesimulation models” represents a grand chal-lenge that tends to inhibit the extensive use ofsimulation in production process design [10].While machine learning approaches have beenimplemented to support production scheduling[18, 26, 21, 14], the integration of data-drivenpredictive models with simulation models israre. Equally lacking is the development ofdata-driven system level models or the imple-mentation of knowledge-based approaches thattake advantage of real-time sensor data to auto-matically generate simulation models and informproduction systems and scheduling [29, 22, 4].

This research explores leveraging upfrontdata analysis and lower complexity simulationmodels to reduce the problem-solving cycle andsupport more expansive studies. In particular,the present paper discusses the integration ofdata-driven predictive and simulation modelsin the context of a partially automated aircraftassembly process with the goal to assess theimpact of disruptions on the production systemand identify how events such as quality assurance(QA) and repair processes can be best handled.

Section 2 describes the system of interest tothis study. Section 3 discusses the approach de-veloped, while Section 4 provides additional con-text to its implementation. Section 5 presentssome results from the simulation study. Finally,Section 6 provide some concluding remarks.

2 System of Interest

This research focuses on a partially automatedassembly process of a large commercial aircraft.The automated drilling systems is composed oftwo types of drillers:

• Drillers A are mounted on rails to sweep

2

Enabling the Digital Factory through the Integration of Data-Driven and Simulation Models

Fig. 1 : Transforming Data into Intelligence

around the circumference of the fuselagejoin, drilling and fastening as they go

• Drillers B move across the side of the fuse-lage, joining the upper shell of the fuselagesection to the lower corresponding section

Both drillers are provided custom drillingprofiles for each hole type and material stack-upconfigurations.

Manual assembly processes are also tak-ing place alongside the drilling tasks withmultiple technicians installing various compo-nents/elements throughout the section of thevehicle.

The data leveraged for this research consists of:

• Automated drilling data for 750,000 hole-by-hole entries for 325 aircraft, including:

– Hole and process information (diam-eter, material, drill only or drill, coun-tersink, fill)

– Process parameters (XYZ hole loca-tion, stack thickness, tool life, etc.)and measured values/process outputs(proved diameter, error messages,etc.)

• Production data, including:

– Actual and scheduled start and endtimestamps for 250 aircraft at the po-sition considered,

– All manual assembly processesalongside drilling tasks

– “Baseline” assembly processes andnon-conformances, quality shakes,emergent removals, etc.

While significant efforts were required toclean and format the data prior to conductingany type of analysis, these efforts are beyond thescope of this paper and thus not discussed herein.

3 Approach

The approach developed focused on integratinginto a digital testbed, predictive data analytics,an hybrid simulation model and process miningheuristics with the goal to provide predictive andimpact assessment capabilities and support moreinformed scheduling decision making.

This approach, illustrated in Figure 2, is bro-ken down into three major tasks:

• Task 1 - Automated Drilling Cycle Timeand Process Analysis: The purpose of thistask is to develop a cycle time predictionmodel for individual drilling job using the

3

OLIVIA J. PINON*, DENNIS JL SIEDLAK*, DIMITRI MAVRIS*

Fig. 2 : Overall Approach

vast amount of automated drilling data col-lected by the system.

• Task 2 - Production Schedule Analy-sis and Baseline Production ScheduleGeneration: The purpose of this task isto leverage process mining heuristics tostochastically generate a baseline produc-tion schedule based on the actual comple-tion of previous aircraft. Such capabilityis particularly important in this context dueto the absence of actual “baseline” from theraw data.

• Task 3 - Process Simulation and ImpactAnalysis: The purpose of this task is todevelop a unified simulation model whichintegrates both a drilling and a schedulemodel. The unified model is an enhancedagent-based model with process miningheuristics added to simulate realistic inter-actions between manual processes (remov-ing rivets, inspections) and automated pro-cesses (drilling holes) within the aircraftassembly line.

3.1 Task 1 - Automated Drilling Cycle Time& Delay Classification and DisruptionAnalysis

The purpose of this task is two-fold: 1) Developdata-driven models to predict the cycle time of in-dividual drilling jobs and, 2) Identify the sourcesof common disruptions and delays.

3.1.1 Predictive Modeling of Drilling CycleTime

Multiple predictive modeling techniques and ap-proaches were investigated. A first approach,which focused on developing an adaptive regres-sion model to predict the cycle time of individualdrilling jobs, was first attempted. This approach,illustrated in Figure 3, integrated Random For-est models for error message predictions withina Neural Network model to predict drilling cycletime.

A second approach focused on develop-ing a multi-layered perceptron (MLP) using theLevenberg-Marquardt algorithm for weight op-timization was then implemented that providedbetter model predictive capabilities. This par-ticular model was trained using the automateddrilling data provided by the system (Figure 4).The model used 144 parameters, including ma-

4

Enabling the Digital Factory through the Integration of Data-Driven and Simulation Models

Fig. 3 : Adaptive Regression Model

chine number, remaining tool life, mechanicalproperties of the processing layers, and machinesettings and had 20 to 30 hidden nodes. It pre-dicted the cycle time of individual drilling jobswith a R2 value of ∼ 0.9 for both training andvalidation samples (Figure 5).

Fig. 4 : Model architecture: the model developedas a similar topology but uses 144 input nodesand 20 ∼ 30 hidden nodes

Fig. 5 : Actual vs. Predicted Cycle Time

3.1.2 Identification of Source of Delays

Gaps between drilling jobs were also studied tofacilitate the modeling of the overall drilling pro-cess. In particular, the following sources of de-lays were identified and integrated in the modeldeveloped as part of Task #3:

• Shift change

5

OLIVIA J. PINON*, DENNIS JL SIEDLAK*, DIMITRI MAVRIS*

• Tool change: either due to a change in pro-cess or a lack of remaining useful tool life

• Work zone interferences: the relative pos-sible collocation of the drillers providesopportunities for disruptions. Work zoneinterferences between Drillers A and B, orbetween two drillers A at the top of thefuselage, for example, are known to cre-ate delays. Similarly, interactions betweendrillers and manual tasks are also cause fordisruptions

• Other delays unexplained by drilling dataalone (part shortages, rework, mainte-nance, etc.)

3.2 Task 2 - Production Schedule Analysisand Baseline Production Schedule Gen-eration

The value of simulation in production analyticsis well understood; however, the time and costto develop a simulation model is sometimes pro-hibitive. This is particularly true when the pro-cess to be simulated lacks a clear precedenceflow, as is the case for the problem at hand.Hence, developing a model to represent such pro-cess and, ultimately provide a testbed to examineimprovements, is extremely challenging.

When faced with a lack of baseline, it is par-ticularly important to identify the overarchinggoal of the simulation model. If the goal is toexamine the logic of how the order of jobs to becompleted is chosen, then a very detailed sim-ulation model may be required. In the contextof this study, where the focus is on identifyinghow events (such as QA and repair processes)impact the overall production flow, the simula-tion model is intended to provide a representativecase to account for these events. The followingthree pieces of data/information are needed forthe simulation:

1. Representative process order

2. Job constraints: jobs that can not beworked on at the same time

3. Job completion times

Identifying and codifying this information iscommonly a very time consuming process whichmakes completing such simulation studies a chal-lenge. Fortunately, there exists a wide varietyof information available from a myriad of datasources throughout the factory. The followingsections discuss how data analytics is leveragedto automatically extract the data required for thesimulation study.

3.2.1 Process Flow Identification

The first piece of information to extract from pro-duction data is a potential process flow. How-ever, unlike well-defined, automated processes,the assembly process of interest does not fol-low a strict precedence network. In other words,the assembly schedule changes from one vehi-cle to the next: in some cases, jobs completednear the beginning of the process are completednear the middle or end on a later vehicle. There-fore, this assembly process does not lend itself tothe use of more traditional process mining tech-niques - techniques that attempt to identify deci-sion branches in processes such as those found incall centers.

Yet, because the simulation only requires arepresentative job schedule (extracted baseline),the flow selected by the algorithm does not needto perfectly conform to process constraints. Toarrive at an average impact of a selected strategy,a number of potential schedules can then be sim-ulated.

The data used to determine the representativeschedule contains the start and end times for ev-ery job competed for a sub-assembly process forapproximately 100 vehicles. Approximately 700jobs are completed per vehicle during this overallprocess.

The resulting schedule needs to be repre-sented as a precedence network that can thenbe executed as a Simio process flow. Simio(SImulation Modeling framework based on In-telligent Objects) [19] was chosen as the model-ing software for this work because 1) its object-oriented nature supports the automated genera-

6

Enabling the Digital Factory through the Integration of Data-Driven and Simulation Models

tion of models, 2) the integration between themodel and data-driven predictive models can beachieved using Simio’s API, and 3) it supportsdistributed computing.

The algorithm discussed below aims at iden-tifying a job ordering as well as precedence re-lationships. As mentioned above, the goal is notto identify the precedence network, since the as-sembly process does not follow a unique one, butrather identify one that is representative enoughof the overall process.

The algorithm operates by analyzing histori-cal sequences to identify the typical sequence po-sition for each job. The first step is to identifythe processes that occur in the majority of the se-quences and their corresponding median processtime and sequence location. These jobs are theninitially ordered based on their median sequencelocation. As a first cut, precedence relations areadded between jobs that are completed before an-other in 95% of the historical schedules provided.For example, if a fitup operation is finished be-fore a drilling operation 95% of the time, then aprecedence relationship is added.

With these “firm” predecessors set, softer,“preference” constraints are added based on aweighted probability draw. This is where thestochasticity of the process comes about. Theprobability that job i is completed before job jis calculated by Equation 1. This is intended tobreak apart jobs that commonly occur close toeach other in the sequence. Therefore, if a joboften occurs before another that is close in thesequence, it has a higher probability of being as-signed a precedence relationship. If a pair failsthe probability check, then no precedence rela-tionship is added.

Once the initial set of precedence relation-ships are identified, the job start and end timesare assigned. The start and end times are enteredby technicians during job execution, so there maybe some variability and errors that must be ac-counted for. The jobs being ordered based ontheir median historical sequence position, thestart times are set as the maximum of their pre-decessors’ end times. As some cycles are gen-erated from the initial precedence identification,

Percent of the time Jobi occurs before Job j

Median sequence position between Job j and Jobi(1)

Process Sequence Preference Probability

this process is repeated multiple times to get arepresentative spread of the start times.

While this provides a good first cut at a rep-resentative sequence, additional data can be con-sidered that add more constraints and eventuallylead to a more representative schedule. First, if ajob has never been completed in parallel with an-other job but appears in parallel in the generatedschedule, then they can be separated by adding aprecedence relationship between them. In suchinstances, jobs are separated such that the jobwith the earlier historical sequence position is ex-ecuted first. Additionally, work zone constraintscan be considered. Hence, if two jobs requireaccess to the same work zone, then they cannotbe completed in parallel and must be separated.The identification of work zones is discussed inthe following section. Following the additionof these new constraints, the process start timesare recalculated and the schedule can be used bySimio.

In summary, the process to generate test pro-duction schedules is as follows:

1. Identify “baseline” processes (those thatoccur in 95% of the provided schedules)

2. Identify a duration for each baseline jobas the median of the previous completiontimes excluding weekends

3. Identify the median sequence position foreach baseline job

4. Order the jobs based on their median se-quence position for the schedule to be gen-erated

5. Identify, for each job, all other jobs that“always” (95% of the time) occur beforethe current. Add precedence relationshipsbetween these jobs to enforce this ordering.

7

OLIVIA J. PINON*, DENNIS JL SIEDLAK*, DIMITRI MAVRIS*

6. Further add “preference” predecessorsbased on Equation 1

7. Assign start and end times based on the de-fined predecessors

8. Separate parallel tasks based on historicaland work zone constraints

9. Recompute start and end times to take intoconsideration the newly added constraints

3.2.2 Work Zone Constraint Identification

During the assembly process, multiple techni-cians are installing various components/elementsthroughout the section of the vehicle. However,jobs that are in the same area of the vehicle can-not be processed at the same time due to physicalconstraints or foreign object debris (FOD) con-siderations. Identifying work zone constraints isthus critical to the modeling exercise as two jobsthat have to be done in the same work zone can-not be executed in parallel.

These work areas, however, are not nicelylaid out in the process information and con-sequently can not be readily integrated inthe model. Fusing information availableabout historical process completions, type ofparts/components to be installed as well as theirphysical locations can help address this issue.Hence, information from the historical processcompletions can be combined with informationabout the parts and components to be installed toget a sense of what those work areas are. Becausejob completion information provides informationabout tasks that have been completed in parallelin the past, it is fair to assume that these tasks donot interfere with each other.

The physical location of the components inthe vehicle and their associated jobs are alsoavailable. Therefore, jobs that have associatedparts can be tagged with a physical location. Thegoal is to then divide the vehicle into work zonesthat can be used to constrain the process.

The work zone separations are identified us-ing an optimization algorithm. Figure 6 shows anexample work zone separation. The current vehi-cle has a temporary floor separating the upper and

paccept = e∆ fT (2)

Probability of Accepting Worse Solution

lower sections, so work can be completed inde-pendently in the upper and lower halves. Withineach upper or lower section, the optimization al-gorithm selects the location of the work zone di-viders. The goal of the optimization algorithm isto minimize the number of jobs that are classifiedin the same work zone and occur in parallel. Inother words, the optimizer helps ensure that jobsthat are being completed in parallel are in differ-ent work zones.

Fig. 6 : Example Work Zone Separators

The optimization is driven by a SimulatedAnnealing (SA) algorithm. The dividers are ini-tially evenly spaced. Then, during each iteration,the algorithm randomly selects a divider to move.The selected divider can move anywhere betweenthe two adjacent dividers with a 5 foot bufferzone to ensure that the created zone is about wideenough for a technician to complete a task. Theobjective function is then evaluated and the newsolution is either accepted or rejected based onEquation 2, where ∆ f is the change in objectivefunction and T is the current annealing tempera-ture. As the problem begins to “cool”, the tem-perature is dropped to reduce the probability ofaccepting a worse solution.

Figure 7 presents an example of the optimiza-tion’s progress. With the regularly spaced di-visions from the first iterations, there are about55,000 overlapping jobs. Following the opti-mization, the algorithm identifies a solution withabout 37,000 overlapping jobs. This solution canthen be used to both separate jobs when gener-

8

Enabling the Digital Factory through the Integration of Data-Driven and Simulation Models

ating the schedule and enforce constraints in thesimulation model.

Hence, with the work zones identified, thegenerated schedule can be updated to observe thework zone constraints (step 8 in the schedule gen-eration process). This is accomplished by check-ing the work zones that each task occupies in theschedule. If tasks that are planned in parallel oc-cupy the same zone, then they are separated byadding a precedence relationship between themand the schedule start times are recalculated.

The simulation model discussed in Section3.3.3 also enforces these constraints during theexecution of a schedule. Each work zone is mod-eled as a seizable resource that is required to pro-cess a task. As such, if a task attempts to start butits required work zone is occupied, then it willwait to start until the work zone becomes avail-able.

3.3 Task 3 - Process Simulation and ImpactAnalysis

The overarching simulation model integrates ascheduling model (generated schedule and iden-tified work zone constraints) and a drilling modelthat includes the individual drilling cycle timepredictive models. As such, the model leveragesthe two sources of data discussed in Section 2:

• Drilling data that is used to generate theprediction models for the driller cycletimes as well as identify the drilling tasks(and drillers) that belong to each individualNC program.

• Production data, which includes locationand time information regarding both man-ual tasks and NC programs, expected andactual durations and other specific taskscharacteristics such as description, type,etc

The characteristics of each model are dis-cussed below.

3.3.1 Scheduling Model Characteristics

The scheduling model combines the input ofa job schedule that is generated stochastically

based on the identified baseline productionschedule discussed in Section 3.2. Hence, themodel stochastically varies the processing time ofthe tasks within the pre-generated baseline pro-duction schedule based on information from realschedules. The model can also evaluate whichtasks can be completed in parallel based on:

• Required predecessors (processes withinthe same line that have to be completed)

• Work zone constraints, which are spatial-dependent, and determine whether jobs canbe completed in parallel or not

The model also has the potential ability to re-schedule tasks based on disruptions happeningwithin the system.

3.3.2 Drilling Model Characteristics

The drilling model simulate sequences of NCprograms and has the following characteristics:

• It integrates the predictive models fordrilling cycle time developed in Task #1

• It allows for work-zones to be simulated.Hence, each driller has working space as-signed with respect to well-defined spacingrules

• It allows for two types of interference to becaptured:

– Work-zone interference: betweenDriller A and Driller B)

– Driller A interference at the top ofthe fuselage (with variable thresholdadded in inches)

• It allows for different interferences rules tobe enforced

– Continue with Line: This model ben-efits the machine that uses repetitivetasks, as long as it was the first onerepeating it, which means that everytime a driller used a resource, and willre-use that resource in the next task,

9

OLIVIA J. PINON*, DENNIS JL SIEDLAK*, DIMITRI MAVRIS*

Fig. 7 : Work Zone Optimization Progress

it will have priority for that resource,halting other possible takers until itfinishes all successions with that par-ticular resource

– First Come First Served (FCFS) :Will assign a resource to any drillerwaiting for one as soon as one is re-leased, no matter if the driller whojust released needs the same resourcefor the next task

– Driller A/Driller B Prioritization:Will give priority to either Driller Aor Driller B

3.3.3 Overarching Model Characteristics

The overarching model simulates realistic in-teractions between manual processes (remov-ing rivets, inspections) and automated processes(drilling holes) within the assembly line consid-ered. As such, the main purpose of this model isto serve as a platform where:

• Disruptions within the schedule involvingmanual and automated tasks can be imple-mented and analyzed to determine their im-pact

• Each of the sub-components can be used inisolation to make more specialized analysiseither at the schedule or the driller level

• More data can be used to simulate morecomplex factory behaviors

The overarching model thus benefits from thecapabilities of both Scheduling and Drilling mod-els. The Scheduling model, instead of deter-mining the tasks times via regressions, generatestasks (NC program sequences) that the Drillingmodel has to complete. The Drilling model thensignals the end of a tasks so that the Schedul-ing model can proceed and carry on with theschedule. Both drilling interferences within theDrilling model and work zone constrains withinthe Scheduling model are taken into account.

3.3.4 Data Flow and Overall Architecture

The data flow, overall architecture and comput-ing languages/platforms leveraged are illustratedin Figure 8. The generated schedule is outputinto a format that can be read into the overar-ching model. The overarching model in Simioevaluates a schedule based on precedence rela-tionships while allowing for stochasticity in the

10

Enabling the Digital Factory through the Integration of Data-Driven and Simulation Models

Fig. 8 : Data Flow, Overall Architecture and Computing Languages/Platforms Leveraged

process time and job success. Simple triangulardistributions are fit to the historical job comple-tion times to provide the process times for themodel. For those jobs that require a work zonebecause they have a component/element/part de-fined, the simulation will seize the correspond-ing work zone resource. In this way, the simula-tion can enforce the work zone constraints if theschedule needs to be modified during the simula-tion run.

4 Implementation

One question driving the development of the sim-ulation model involved analyzing the impact of ajob failure and subsequent non-conformance re-pair on the overall process. Figure 9 presents anoverview of the simulation flow when a QA fail-ure is encountered. In this figure, Job 231 failsQA and must go through an engineering analy-sis and repair process. While this is ongoing, Job232 is required to wait for 231 to be reworked,thereby delaying the process. However, duringthis time, other jobs that are planned later in theschedule (302, 305, 315, and 325 in this instance)can be completed as they are not tied to Job 231’sprecedence relationship. As such, the simulationdecides to complete these tasks while the reworkon 231 is being completed to help minimize theimpact of that failure.

When investigating the impact of rework andalternative mitigation strategies, the two researchquestions of interest are:

1. Research Question #1: How should I pri-oritize the list of quality problem fixes thatmust pass through engineering before be-ing resolved?

2. Research Question #2: When a disruptionoccurs, what strategies best mitigate theimpact of the disruption?

Research question #1 focuses on identifyinghow quality issues should be prioritized throughnon-conformance review. There are only somany engineers who can recommend and ap-prove non-conformance fixes, so it is possiblethat quality issues can build up and lead to sig-nificant delays in the process. As such, differentranking heuristics can be identified to best mit-igate these problems. The data about the non-conformance review was not available at the timeof this study, so this question will be investigatedin future studies.

Research question #2 focuses on finding howvarious strategies to select tasks to completewhile waiting for the QA analysis impact theoverall completion time of the schedule. As such,various strategies are investigated including first

11

OLIVIA J. PINON*, DENNIS JL SIEDLAK*, DIMITRI MAVRIS*

Fig. 9 : Simulation Schedule Recovery Process

in/first out, shortest available job, longest avail-able job, and processing the job with the mostsuccessors first.

For both research questions, the primary met-ric of interest is the process time required to com-plete the schedule. The simulation is run for mul-tiple replications (because the process times andQA failures are stochastic) and for multiple test-ing schedules (because the schedule generationitself is also stochastic). To demonstrate the anal-ysis provided by this approach, sample resultsfrom these studies are presented in the followingsection.

5 Results

The simulation results presented in this sectioninvestigate different strategies to employ to re-cover from a non-conformance. Using the pro-cess described in Figure 2, simulation experi-ments are conducted to test the following strate-gies to select tasks to complete while waiting fora non-conformance inspection to be completed:

• Average processing time/schedule loca-tion: This rule favors completing longertasks that are earlier in the schedule in anattempt to free resources for smaller jobsonce the non-conformance review is com-pleted

• Earliest in schedule: Favors tasks purelybased on when they happen in the sched-ule to try to complete predecessors for asmany upcoming tasks as possible

• Largest average duration: Favors longtasks to free resources later on

• Smallest average duration: Favors shortertasks to complete as many as possible dur-ing the rework cycle

• Most successors: Favors tasks that have themost successors to open up more possibili-ties to complete future tasks

Upon encountering a non-conformance, thesimulation uses the rule selected for that runto choose task(s) to complete during the non-conformance review. Figure 10 shows the impactof the selection rule choice on the overall processflow time. This figure is presented as a SimioMeasure of Risk & Error (SMORE) plot. This issimilar to a standard box-and-whiskers plot ex-cept that the brown and light blue boxes centeredaround the mean and quantiles denote the statis-tical confidence of those measures. The lighterblue bars extending above the whiskers show thehistogram of raw results. From Figure 10, thetwo strategies targeting longer duration tasks forcompletion take longer to finish the overall pro-cess than the other three strategies. Figure 11,which shows the number of tasks completed dur-ing the rework process, provides a possible ex-planation for this impact. By targeting long du-ration tasks, fewer tasks can be completed dur-ing the rework process (as expected). Hence, itappears that this overall process benefits morefrom completing more short duration tasks than

12

Enabling the Digital Factory through the Integration of Data-Driven and Simulation Models

Fig. 10 : Impact of Selection Rule on Overall Process Flow Time

Fig. 11 : Impact of Selection Rule on the Number of Tasks Completed During Rework Periods

fewer, long duration tasks. However, selectingto complete the longest duration tasks first doesproduce the smallest maximum processing time.Consequently, there may exist a hybrid strategythat would allow to complete many short duration

tasks while preventing having long tasks near theend of the process.

These results demonstrate the nature of theinformation that can be generated by integrat-ing multiple data sources, data analytics, machine

13

OLIVIA J. PINON*, DENNIS JL SIEDLAK*, DIMITRI MAVRIS*

learning, process mining techniques, and simula-tion. While not mature enough to be used op-erationally, this process is intended to provide astarting point for future, more focused analyticsand simulation studies.

6 Concluding Remarks

The present paper discussed the integration ofdata-driven and simulation models in the contextof a partially automated aircraft assembly processto 1) assess the impact of job failure and subse-quent non-conformance repair on the overall pro-cess, and 2) test strategies to best recover fromsuch disruptions. In particular, this paper has dis-cussed and/or demonstrated:

• The utilization of statistical methods andprediction techniques to identify predictivefeatures to be included in the cycle timeprediction model

• The investigation of interaction betweenboth automated and manual jobs to identifyrelationships and disruptions to the produc-tion system

• The creation of a data-driven, robust sched-ule analysis tool to identify most com-monly occurring sequences of jobs fromactual schedule completion information

• The creation of a work zone identificationtool to infer spatial constraints from sched-ule completion information as well as partinstallation requirements

• The integration of both data-driven andsimulation models to:

– Evaluate critical tasks that signifi-cantly impact the overall flow

– Identify promising recovery rulesand/or schedules

While not mature enough to be used on theshop floor, this study has nonetheless contributed

to demonstrate the value of integrating data-driven and simulation models as a means to ob-tain more visibility on how to best recover fromdisruptions.

7 Acknowledgments

The authors would like to acknowledge Boe-ing Research & Technology for funding this re-search. In particular, the authors would like tothank Jonathan Fulton for sharing his expertiseand providing support and feedback during thedevelopment of the approach presented herein.Finally, the authors would like to acknowledgeDr. WoongJe Sung, David Solano and BurakBagdatli for their significant contributions to thiswork.

8 Copyright Statement

The authors confirm that they, and/or their company ororganization, hold copyright on all of the original ma-terial included in this paper. The authors also confirmthat they have obtained permission, from the copy-right holder of any third party material included in thispaper, to publish it as part of their paper. The authorsconfirm that they give permission, or have obtainedpermission from the copyright holder of this paper, forthe publication and distribution of this paper as part ofthe ICAS 2018 proceedings or as individual off-printsfrom the proceedings.

References

[1] Abellan-Nebot J and Subiron F. A review ofmachining monitoring systems based on artifi-cial intelligence process models. InternationalJournal of Advanced Manufacturing Technol-ogy, Vol. 47, pp 237–257, 2010.

[2] Alaeddini A and Dogan I. Using bayesian net-works for root cause analysis in statistical pro-cess control. Expert Systems with Applications,Vol. 38, pp 11230–11243, 2011.

[3] Baur C and Wee D. Manufacturing’s next act.Technical report, McKinsey Digital, 2015.

[4] Bergmann S and Strassburger S. Challenges forthe automatic generation of simulation models

14

Enabling the Digital Factory through the Integration of Data-Driven and Simulation Models

for production systems. Proc SCSC ’10 Pro-ceedings of the 2010 Summer Computer Simula-tion Conference, Ottawa, Ontario, Canada, July2010. Society for Computer Simulation Interna-tional.

[5] Choudhary A, Harding J, and M.K. T. Data min-ing in manufacturing: A review based on thekind of knowledge. Journal of Intelligent Man-ufacturing, Vol. 20, pp 501–521, 2009.

[6] Davis J, Thomas E, Graybill R, Korambath P,Schott B, D, Wang J, and Wetzel J. Smartmanufacturing. Annual Review of Chemical andBiomolecular Engnieering, Vol. 6, No 141-160,2015.

[7] Dey S and Stori J. A bayesian network approachto root cause diagnosis of process variations. In-ternational Journal of Machine Tools & Manu-facture, Vol. 45, pp 75–91, 2005.

[8] ElMaraghy H and ElMaraghy W. Smart adapt-able assembly systems. Proc Procedia CIRP 44,2016.

[9] Elovici Y and Braha D. A decision theoreticapproach to data mining. IEEE Transactions onSystems, Manufacturing, and Cybernetics, PartA, Systems and Humans, Vol. 33, No 1, pp 42–51, 2003.

[10] Fowler J. W and Rose O. Grand challengesin modeling and simulation of complex manu-facturing systems. Simulation, Vol. 80, No 9,pp 469–476, 2004.

[11] Guh R.-S. Integrating artificial intelligence intoon-line statistical process control. Quality andReliability Engineering International, Vol. 19,pp 1–20, 2003.

[12] Harding J. A, Shahbaz M, Srinivas, and Ku-siak A. Data mining in manufacturing: A re-view. Journal of Manufacturing Science andEngineering, Vol. 128, No 4, pp 969–976, 2006.

[13] Hartmann B, King W. P, and Narayanan. Digitalmanufacturing: The revolution will be virtual-ized. Technical report, McKinsey Digital, 2015.

[14] Huet A. Optimization and analysis aid via data-mining for simulated production systems. Eu-ropean Journal of Operational Research, Vol.173, pp 827–838, 2006.

[15] Kelly R. The digital-manufacturing revolution:How it could unfold. Technical report, McKin-sey Digital, 2015.

[16] Lee J, Lapira E, Bagher B, and Kao H.-a. Recentadvances and trends in predictive manufacturingsystems in big data environment. Manufactur-ing Letters, Vol. 1, pp 38–41, 2013.

[17] Lee J, Lapira E, Yang S, and Kao H.-A. Pre-dictive manufacturing system - trends of next-generation production systems. Proc 11th In-ternational Federation of Automatic Control(IFAC) Workshop on Intelligent ManufacturingSystems, Sao Paulo, Brazil., May 2013.

[18] Li X and Olafsson S. Discovering dispatchingrules using data mining. Journal of Scheduling,Vol. 8, pp 515–527, 2005.

[19] LLC S. Simio, version 8.141, 2016. Sewickley,Pennsylvania.

[20] Pham D and Afify A. Machine learning tech-niques and their applications in manufacturing.Part B: Journal of Engineering Manufacture,Vol. 219, pp 395–412, 2005.

[21] Priore P, Gomew A, Pino R, and Rosillo R.Dynamic scheduling of manufacturing systemsusing machine learning: An updated review.Artificial Intelligence for Engineering Design,Analysis and Manufacturing, Vol. 28, pp 83–97,February 2014.

[22] Soderberg R, Warmefjord K, Carlson J. S, andLinkdkvist L. Toward a digital twin for real-time gepmetry assurance in individualized pro-duction. CIRP Annals - Manufacturing Technol-ogy, Vol. 66, pp 137–140, 2017.

[23] Tambe P. Big data investment, skills and firmvalue. Management Science, Vol. 60, No 6,2014.

[24] Uhlemann T. H.-J, Schock C, Lehmann C,Freiberger S, and Steinhilper R. The digitaltwin: Demonstrating the potential of real timedata acquisition in production systems. ProcProcedia Manufacturing - 7th Conference onLearning Factories, CLF 2017, Vol. 9, pp 113–120, 2017.

[25] Weese M, Martinez W, Megahed F. M, and Al-lison J.-F. L. Statistical learning methods ap-plied to process monitoring: An overview andperspective. Journal of Quality Technology,Vol. 48, No 1, pp 4–27, 2016.

[26] Witkowski T. Data-based scheduling produc-tion with simulation and optimization methods.Proc 7th IEEE International Conference on In-

15

OLIVIA J. PINON*, DENNIS JL SIEDLAK*, DIMITRI MAVRIS*

telligent Data Acquisition and Advanced Com-puting Systems: Technology and Applications,2013.

[27] Wuest T, Weimer D, Irgens C, and Thoben K.-D. Machine learning in manufacturing: Ad-vantages, challenges, and application. Produc-tion & Manufacturing Research, Vol. 4, No 1,pp 23–45, 2016.

[28] Zoeller K. Digital factory is key to increasingmanufacturing capacity say industrial automa-tion north america 2016 exhibitors. IndustrialTechnology, 2016.

[29] Zou J, Chang Q, Lei Y, and Arinez J. Productionsystem performance identification using sensordata. IEEE Transactions on Systems, Man, andCybernetics: Systems, Vol. 48, No 2, pp 255–264, February 2018.

16