Towards Automated Parameter Optimisation of Machinery

by Persisting Expert Knowledge

Richard Nordsieck

, Michael Heider

, Andreas Angerer

and J

org H

ahner

XITASO GmbH IT & Software Solutions, Augsburg, Germany

Organic Computing Group, University of Augsburg, Augsburg, Germany

Keywords: Additive Manufacturing, Transfer Learning, Domain Adaption, Machine Learning, Knowledge Representa-

tion.

Abstract:

Commissioning of machines takes up a considerable share of time and money of the total cost of developing

a machine. Our project aims at developing an approach to decrease the time needed to commission machines

by automating parameter optimisation with the help of formalised expert knowledge. The approach will be

developed on the Fused Deposition Modelling (FDM) process, which is an additive manufacturing technique.

We pay particular attention to keeping the approach sufﬁciently abstract to be applied to machines from other

domains to beneﬁt its industrial application.

1 INTRODUCTION

Commissioning of machinery can consume up to 25%

of the total project duration according to a study of the

German Machine Tool Builders’ Association (VDW,

1997). They also state that up to 90% of this time

is spent on commissioning the electronics and control

software. As stated in (Abel, 2017), the commission-

ing process is done mainly by specialized workers, as

it is a complex process and needs expert knowledge.

The geographic location of manufacturers of ma-

chinery and their customers is not necessarily identi-

cal. This leads to employment of highly-skilled pro-

fessionals abroad, which is even more cost intensive

due to increased management costs as well as travel

and accommodation expenses. Therefore, commis-

sioning is a prime target for cost and time optimisa-

tion of the manufacturing process overall, which is

crucial for manufacturers to compete in a globalised

economy. Furthermore, customers of manufacturing

systems also have a strong interest in short commis-

sioning times, as they are a prerequisite for expedi-

tious use of newly acquired systems. Given the trend

towards Reconﬁgurable Manufacturing Systems (Ko-

ren, 2006), shorter commissioning times gain even

more importance: Every reconﬁguration of a manu-

facturing system requires a subsequent commission-

ing to ensure proper operation of the machinery in

changed conﬁguration.

We differentiate between static and dynamic com-

missioning. Static commissioning denotes the me-

chanical installation, operative checks of a machine

or plant and, if needed, programming of control soft-

ware. After static commissioning, single units and

their integration can be successfully tested. This can

usually be accomplished off-site e.g. at the man-

ufacturer, at least for initial machine setup. Dy-

namic commissioning, however, describes the ﬁne-

tuning of process parameters until a satisfactory qual-

ity level of the produced part is achieved. In contrast

to static commissioning, it can be seen as proof of

a correctly working system under environmental in-

ﬂuences. During dynamic commissioning, changes

in the parametrisation of machinery as well as slight

modiﬁcations to the control software might be neces-

sary.

To reduce the duration of commissioning, the con-

cept of virtual commissioning was introduced (Lee

and Park, 2014; Hoffmann et al., 2010). In this

concept, commissioning is simulated in varying lev-

els of detail, which allows testing the interactions of

components as well as the control programs, which

are usually based on programmable logic controllers

(PLCs)). This allows for early detection and remedy-

ing of faults, which account for a substantial amount

of time spent during the commissioning procedure

(VDW, 1997). Thereby, virtual commissioning is

able to reduce the static part of the commissioning

to the mechanical and electrical installation. How-

ever, one drawback of virtual commissioning is the

406

Nordsieck, R., Heider, M., Angerer, A. and Hähner, J.

Towards Automated Parameter Optimisation of Machinery by Persisting Expert Knowledge.

DOI: 10.5220/0007953204060413

In Proceedings of the 16th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2019), pages 406-413

ISBN: 978-989-758-380-3

Figure 1: Schematic visualisation of machine lifetime. First

the machine is constructed, which is followed by com-

missioning. Later on in the lifetime of a machine re-

parametrisation might be needed. While the top bar shows

a traditional lifetime, the middle depicts a lifetime, which

is optimised by the use of virtual commissioning (VC). The

bottom bar shows the lifetime after automatic parameter op-

timisation.

effort required to create a digital twin of the plant or

machine. While this may amortise quickly for larger

manufacturing systems, plants or manufacturers that

sell the same machine multiple times, manufacturers

of special-purpose machinery ﬁnd it hard to justify the

recurring expenditure.

Nevertheless, the actual commissioning, which in-

cludes the dynamic commissioning, remains neces-

sary as it is partly depending on environment-speciﬁc

conditions (Abel, 2017). Thus, adjustments that com-

pensate inﬂuences e.g. by temperature, humidity or

raw material composition can only be performed on-

site where the machine is operated. Note again that

this can be necessary during initial on-site installa-

tion of machinery, but also after re-conﬁgurations due

to changes in production processes or target prod-

ucts. Consequently, our approach aims at reducing

the time required for the dynamic part of commis-

sioning shown in orange horizontal lines in Figure

1, that is parameter optimisation. It can be applied

in conjunction with virtual commissioning, which re-

duces static parts of the commissioning process, or as

a stand alone solution.

Expert knowledge required for parameter opti-

misation is traditionally acquired by working under

the supervision of more experienced commissioners

in a trainer-trainee fashion as well as commission-

ing a vast amount of machines. Due to demographic

change and subsequent shifts in the workforce many

experts will retire in the foreseeable future. Manu-

facturers are therefore posed with the challenge of

how to preserve their knowledge and stay competi-

tive. While commissioners might have documented

commissioning procedures in varying detail and for-

mats, few manufacturers have successfully worked to-

wards a comprehensive knowledge base, according to

our knowledge.

Similarly to other production processes, Fused

Deposition Modelling (FDM) requires experts to con-

duct parameter optimisations to produce parts of satis-

factory quality. Due to the small form-factor of FDM

machines, it is feasible to control environmental in-

ﬂuences. Furthermore, parametrisations can be tested

on multiple different parts without the need to man-

ufacture tools, as is the case in other production pro-

cesses. Therefore, we chose the FDM process to de-

velop an approach to automatic parameter optimisa-

tion in a controlled environment. We see FDM as a

representative of other production processes that de-

pend on environmental conditions. After a success-

ful veriﬁcation of our approach on FDM, we study

the transferability of the approach to other production

processes, working towards our goal of decreasing

their commissioning times by automating parameter

optimisation.

This paper identiﬁes challenges for automating

parameter optimisation and outlines an approach on

how to solve these. Section 2 provides a detailed

view on production processes of additive manufactur-

ing, based on which we will develop and verify our

approach. Challenges for automating dynamic com-

missioning are introduced in Section 3. Our approach

to automate parameter optimization by combining ex-

pert knowledge and machine learning is detailed in

Section 4. Section 5 concludes this paper.

2 CASE STUDY

As described in Section 1, we decided to develop our

approach on the Fused Deposition Modelling (FDM)

process, which is an additive manufacturing tech-

nique, in which a nozzle lays down material layer-

wise until the desired object is constructed (Gibson

et al., 2015). To facilitate this, the material is heated

above its melting point and extruded through the noz-

zle. While extruding the nozzle moves along a pre-

planned path at a given speed. When the nozzle needs

to move without extruding, e.g. when printing two

non-connected models at once, the material is ﬁrst re-

tracted to decrease pressure in the extruder. Many

printers use a heated bed on which the ﬁrst layer is

placed. The heating increases adhesion and slows

down the effects of thermal expansion (or in this case

contraction). We chose FDM since it is a controllable

manufacturing process with multitudes of inﬂuence

factors and conﬁgurable parameters. In addition, the

results can be analysed by hand without the need of

extensive training, a laboratory or other highly spe-

cialized equipment.

Additive manufacturing has become increasingly

popular since its commercial introduction in the late

Towards Automated Parameter Optimisation of Machinery by Persisting Expert Knowledge

407

Figure 2: Cause and effect (Ishikawa) diagram of the for

this work most important inﬂuence factors on part quality

in an FDM production process.

1980s and early 1990s (Wohlers and Gornet, 2014).

It is mostly used for rapid prototyping and manufac-

turing of small quantities of items, where moulding,

casting and other conventional techniques would be

too expensive. Another beneﬁt of additive manufac-

turing is the ability to produce complex forms. These

can be found in medical applications or if weight is of

concern.

The main development in the scope of this project

will be done on different cartesian and delta FDM

printers using different plastics, such as PLA, ABS

and PETG. After thoroughly evaluating our approach

on multiple FDM printers, we will move on to differ-

ent additive manufacturing techniques such as binder

jetting (BJ) and selective laser sintering (SLS). BJ is a

process, where powdered material is fused into form

by applying a binding agent. SLS is another powder

bed based process, where material is sintered locally

and thereby fused into form. For each process ex-

pert knowledge will be collected and formalised sepa-

rately. While certain expertise can be directly applied

to the other additive manufacturing techniques, some

knowledge is most likely very speciﬁc and differs be-

tween them. Thus, we will be able to test the transfer

of knowledge between different machines of the same

production process as well as between different addi-

tive manufacturing production processes. This marks

the ﬁrst step towards an abstract universal architec-

ture.

The ﬁnal quality of a 3D printed part is depen-

dent on various factors with differing degrees of in-

ﬂuence (Mohamed et al., 2015). Some factors can be

quantiﬁed or measured better than others. The anal-

ysis of those factors can be done in discretionary de-

tail. We have isolated the factors deemed most im-

portant in this context for the case of a generic FDM

process in Figure 2. As we focus on developing an

approach for parameters that are adjustable by soft-

ware, some machine parameters, like nozzle diame-

ter or ﬁlament width, can be seen as set. To offset

the inﬂuences of wear and tear in mechanical parts

on our machine learning systems, we plan on replac-

ing the printers with new off the shelf versions during

the project. We will focus on the commissioning and

initial parametrisation of a system and thus not inves-

tigate machine degradation inﬂuences on optimal pa-

rameter selection.

There are several environmental factors that affect

the quality of a produced part. In the following we

will provide an overview of the environmental factors

most relevant to FDM printing as well as to parame-

ters used to mitigate their effects. Humidity has a sig-

niﬁcant impact on material quality. Due to the time

it requires to infuse the ﬁlament it is more relevant

during storage of ﬁlaments than during the printing

process. Dust can, depending on amount and compo-

sition, decrease print quality and do serious damage

to the machine by inﬂaming. Temperature is another

relevant factor. On a process-agnostic perspective ex-

cessive heat or cold could do substantial damage to

machine parts. On a process-related perspective the

effects are more subtle. When molten plastics cool

down they solidify and contract, the degree of which

is dependent on the temperature and speciﬁc plastic.

In the case of FDM printers this can lead to warping.

Warping is a process, where segments of the printed

part deform (curling upwards, away from the build

plate, along the parts edges). It typically occurs as

the lower-layers of the part cool down too fast. Since

freshly extruded material needs to cool quicker than

it would under room temperature (21

◦

C) most print-

ers include a cooling fan to increase print quality. We

will initially focus on temperature as a factor, as it,

opposed to humidity, directly inﬂuences the process

of FDM printing. Another relevant factor that of-

ten varies under real-word conditions is raw-material

quality and composition, which inﬂuences print qual-

ity due to different temperature and form related prop-

erties. As such, we will conduct experiments, vary-

ing both temperature and raw-materials leading to a

dataset, which we will use to train algorithmic ap-

proaches to parameter optimisation.

The main FDM parameters modiﬁable through

software are either related to temperature, movement

or the preprocessing (i.e. slicing, the generation of

support structures and machine path planning). Print-

ing bed and extrusion temperature can offset low sur-

rounding temperatures while cooling fan speed can

aid the cooling of freshly extruded ﬁlament. Layer

thickness not only inﬂuences the optical appearance

of a part but also the material properties during and af-

ter printing. Bridging is a part of the part construction

process, where the printer tries to extrude material in

an area, where no (immediate) lower layer is present,

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

408

thus bridging two segments of lower layer together.

Bridging becomes harder to accomplish if the part

cools down too fast or too slow. Low speed can also

result in bridges failing. Intelligent support placing

could offset arising issues as bridges would become

shorter. In the best case issues could also be prevented

by simple print orientation changes, e.g. bottom up

(Kozior and Kundera, 2017). Since we want to focus

on parameters that are transferable to other production

processes, we will focus on optimising parameters re-

lating to the actual process as opposed to parameters

changing the structure of the produced part. Vibra-

tions that increase with velocity changes of the nozzle

disrupt precise printing. Lowering the print speed and

accelerations can accommodate for that (Alafaghani

et al., 2017). Slower printing also tends to allow the

extruded material to better set in place, rather than

getting dragged along. Adjusting print speed and ac-

celeration quickly becomes a trade-off between quan-

tity and quality of produced parts. The analysis of our

experimental results will help to quantify this trade-

off.

3 CHALLENGES

Our research hypothesis is that parameter optimisa-

tion has a signiﬁcant effect on the quality of produced

parts and thereby the amount of time required for

commissioning. Consequently, automating parame-

ter optimisation has the potential to effectively reduce

cost and time spent on commissioning manufacturing

systems. To conﬁrm this hypothesis we plan to con-

trast times required for manual parameter optimisa-

tion with those resulting from the application of our

optimisation algorithm. Actually developing a de-

pendable approach for this task, however, poses sev-

eral challenges as outlined in this chapter.

3.1 Identiﬁcation and Quantiﬁcation of

Parameters

In Section 2 we introduced several environmental in-

ﬂuences that can have a detrimental effect on the pro-

duction process of the machine that is to be commis-

sioned. Correctly identifying and quantifying them

is a challenging task. This becomes especially evi-

dent if one considers diverse raw materials and their

possibly differing qualities, e.g. manufacturer or even

batch dependent.

Furthermore, we introduced process parameters

that can be varied to mitigate the detrimental effect

of aforementioned environmental parameters. To val-

idate that these parameters are actually relevant to

the production process and all relevant parameters are

taken into consideration, we will conduct structured

interviews with companies that employ additive man-

ufacturing at an industrial scale and other experts in

the ﬁeld.

Section 2 is limited to processes encountered in

the domain of additive manufacturing. Since we aim

for an approach that can also be applied to other

production processes we need to arrive at a domain-

independent deﬁnition of commissioning. This deﬁ-

nition can then be used to deﬁne similarities between

production processes and their commissioning proce-

dures. To arrive at this deﬁnition we will also conduct

structured interviews over a broad range of producers

of manufacturing systems or manufacturers.

3.2 Data-related Challenges

Commissioning in full, is usually executed before the

handover of the manufacturing system. Parameter op-

timisation can also be required after a change in raw-

material quality, a reﬁt to produce different parts or

extensive maintenance work. Compared to the overall

runtime of a manufacturing system, this leads to very

few opportunities to collect data. To generate data a

simulation of the production process in question could

be employed. However, many production processes

are heavily inﬂuenced by advanced thermodynamics

and material properties. Therefore, simulating them

to the degree of detail that is needed requires large

amounts of time and resources. Furthermore, the

simulation would be limited to one production pro-

cess and not easily adaptable to additional processes,

which would render it uneconomic. A more promis-

ing solution to the data sparsity could be achieved by

incorporating data collected by predictive quality ap-

proaches already employed on similar machines in

the ﬁeld. This would make it easier to empirically

detect correlations between environmental inﬂuences

and corresponding parameter conﬁgurations.

Automatically assessing the quality of chosen pa-

rameters via assessing the quality of the produced

product is another data-related challenge. This would

require the presence of automatic quality measure-

ment systems, which is not commonplace. There-

fore, our envisioned approach should actively request

feedback of experts that manually assess the product’s

quality.

Towards Automated Parameter Optimisation of Machinery by Persisting Expert Knowledge

409

Figure 3: Schematic description of knowledge based (top-

down) and learning systems (bottom-up). Our proposed

approach (ACPE) combines aspects of both and is conse-

quently shown in the middle.

4 TOWARDS AUTOMATIC

COMMISSIONING

As outlined in Section 3, one of the main challenges

for applying learning systems is the scarcity of com-

missionings of identical machines, which leads to few

opportunities for collecting data. To mitigate the re-

sulting uncertainties and training difﬁculties we pro-

pose an approach that combines learning and knowl-

edge based methods. Based on a large amount of

observations, learning systems generalize knowledge.

As such, they work in a bottom–up manner. Contrary

to that, knowledge based systems work as a top–down

approach, inferring facts from a knowledge- and rule-

base (Liao, 2005). A high level overview of these

paradigms and their proposed interactions with our

approach can be seen in Figure 3. This sections out-

lines our approach to achieve automatic parameter op-

timisation for commissioning under the constraint of

sparse data and places it in academic context.

4.1 Learning Quality Prediction and

Strategy Selection

In the context of FDM printing we see machine learn-

ing as a viable approach for achieving suitable quality

prediction and strategy selection. In a reinforcement

learning context a strategy would also be referred to

as a policy. The ideal strategy would be to select pro-

cess parameters (the action the system should take)

for each set of 3D model, printer, environment and

material (forming the state the system ﬁnds itself in)

in a manner that facilitates the highest possible qual-

ity. Predicting the print quality requires an algorithm

to predetermine the quality of a part produced by ap-

plying an action to a speciﬁc state. Quality predic-

tion and strategy selection are generally different ma-

chine learning tasks, namely regression and planning.

We will solve those with different base algorithms

forming an actor-critic method (Sutton et al., 2000).

Nonetheless, the results obtained by quality predic-

tion can greatly improve strategy selection. Given an

advanced quality prediction system each state-action

pair can be evaluated on the ﬂy without the need to

real world test it by printing it, resulting in model-

based learning, which would be considerably faster.

If the described ideal strategy was known, it would

allow instantaneous commissioning of all printers for

all tasks without the need of any commissioning time.

An ideal partial strategy selects optimal actions (re-

turning the highest quality part) for one given printer.

We expect partial strategies to be related to each other

and in conjunction to be able to closely describe the

ideal strategy, if enough partial strategies are known.

Finding a quality prediction system, that is able to

successfully generate reliable quality predictions for

all state-action pairs, will not be achievable due to

the large variety of printers and models. However we

will create partial prediction systems that are able to

perform for individual or small sets of printers, while

trying to maintain an as large as possible variety of

predictable models. Those partial prediction systems

will be able to generalise for similar printers. The

main challenge for constructing the prediction sys-

tem will be the availability of data as there are—to

our knowledge—no public comprehensive 3D print-

ing data sets and as such all training data will have to

be generated. To solve the quality prediction problem

we will investigate the use of various classiﬁcation

techniques such as linear regression, decision trees

(Kotsiantis, 2013), support vector machines and ar-

tiﬁcial neural networks. Depending on the algorithm

we will either model classiﬁcations as individual re-

gression problems or engineer a metric to transform

class based assessments into a single variable descrip-

tion. This is done to ensure comparability. Beyond

the algorithm itself we see two general approaches to

the problem: To either try to directly incorporate the

model as a whole into the classiﬁer or to only use an

abstract representation for it. The former would re-

sult in the use of learned convolutions (in the form of

neural network layers) to extract core features and the

latter would require an advanced (computable) met-

ric to describe model complexity. During this project

we will research the most feasible classiﬁcation al-

gorithm and approach for our problem as they both

have advantages and drawbacks. Convolutional net-

works typically need numerous training samples, in-

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

410

creasing with depth, until good results are achieved.

When, like in this case, the exact targets are un-

known, but rather represent a partial input for classiﬁ-

cation, learning the correct convolution becomes even

harder. Yet they are able to learn complex connections

for which ﬁnding an accurate algorithmic description

would normally require large amounts of time.

To ﬁnd reasonable strategies we plan to utilise Ex-

tended Classiﬁer Systems (XCS) (Wilson, 1995; Butz

and Wilson, 2002) a derivative of Learning Classiﬁer

Systems (LCS) (Holland, 1976). While standard XCS

typically takes binary inputs we will encounter con-

tinuous (e.g. temperatures or speeds) or other non bi-

nary inputs and therefore need the ability to take real

value inputs, which becomes possible in the XCS ex-

tension XCSR (Wilson, 2000). XCSR uses IF-THEN

rules with an assigned quality (accuracy in predicting

reward). Those IF-THEN rules are typically gener-

ated and optimised by a steady-state niche genetic al-

gorithm (GA) but they can also be created by humans,

for example by formalising the knowledge of a pool

of experts as described in Section 4.2. The strategy

itself will need to be able to extrapolate knowledge

from known state-action pairs and combine it with ex-

isting knowledge. To ﬁll existing knowledge gaps, as

formerly deﬁned by Stein et al. (Stein et al., 2018),

active learning (Cohn et al., 1994) can be employed

(Stein et al., 2017a). In addition we will expect to be

able to use interpolation between known classiﬁers to

gain knowledge on inter-laying classiﬁers (Stein et al.,

2016; Stein et al., 2017b). Figure 4 illustrates the ba-

sic procedure. Known points (sets of parameters) in

the multi-dimensional parameter space, that are simi-

lar to the parameter combination for which the quality

is to be determined, are selected and interpolated to

determine the new quality.

4.2 Incorporating Expert Knowledge

Historically, there have been multiple options of inte-

grating expert knowledge with algorithms. The most

prominent are expert systems. Expert systems con-

sist of manually curated knowledge bases that contain

facts and rules, as well as inference engines that de-

duce facts based on the associated knowledge bases

(Liao, 2005). However, formalising expert knowl-

edge to a sufﬁcient degree is a complex task, espe-

cially since experts tend to take a certain amount of

common-sense for granted, omitting crucial informa-

tion about the process. Furthermore they can only

dedicate a limited amount of time since they are con-

stantly required to work on projects creating more di-

rect revenue.

Embeddings of knowledge bases are successfully

Figure 4: Illustration of a three dimensional subset of

the parameter space, that includes the parameters, part to

be produced (the pre-processed 3D Model) and machine

type and the parameter group of environmental inﬂuences.

Points for which the quality is known are displayed as grey

circles. Points for which no knowledge is available are

displayed as orange triangles. Using similarities between

known and unknown points a set of known points (grey

cube) is selected. A function

f allows interpolation of

knowledge for the unknown point.

applied to increase performance of learning systems,

for example in recommender systems (Wang et al.,

2018; Sun et al., 2018) and text understanding (Wang

et al., 2017; Yang and Mitchell, 2019). In the context

of LCS Urbanowicz et al. experimented with the in-

corporation of expert knowledge for guiding the dis-

covery components to explore attribute values found

most predictive (Urbanowicz et al., 2012). In addi-

tion, reasoning tasks have been shown to fall short be-

cause of a lack of formalised common-sense knowl-

edge being present in the data they have been trained

on (Tandon et al., 2018). Similarly, there are ap-

proaches to mitigate this by incorporating formali-

sations of common-sense knowledge in the training

process (Ma et al., 2018). Inspired by the success-

ful combination of knowledge bases and learning sys-

tems outlined above, our approach seeks to mitigate

the effect of scarce training data, by fusing of knowl-

edge based and learning systems.

Knowledge about production processes, relevant

parameters and inﬂuences is readily available in the

industry. However, most of the time it is not docu-

mented in a structured and digitalized way. Therefore,

we propose to work towards a knowledge base by per-

sisting expert knowledge. As a ﬁrst step we will form

a semi-structured natural language knowledge base by

recording expert knowledge of the FDM process with

a markup language. This has the beneﬁt of provid-

ing an easy to use interface for experts. Furthermore,

if embedded in a documentation infrastructure, that

Towards Automated Parameter Optimisation of Machinery by Persisting Expert Knowledge

411

other experts have access to, it provides a way of shar-

ing knowledge throughout the company. While this

infrastructure can already be considered a knowledge

base, it is not yet easily computationally interpretable.

Therefore, it is converted into a more formal represen-

tation consisting of conditions (Lehmann and Magi-

dor, 1992) and triples, that describe relationships be-

tween entities (Nickel et al., 2016). Based on formal

representations, multiple approaches exist in litera-

ture for knowledge base completion by automatically

detecting and creating missing relations (Socher et al.,

2013; Nguyen et al., 2018; Lin et al., 2015; Das et al.,

2017). We will evaluate their applicability to process-

oriented knowledge bases. Lastly, these completed

formal representations are then transformed into em-

beddings which can be incorporated into the learning

algorithms to give them an understanding of process

knowledge. We assume that this has a positive effect

on the amount of data required for training.

As another approach addressing the scarcity of

training data we want to evaluate transformations

between knowledge representations. Knowledge in

LCSs is represented by a learnt set of rules, whereas in

ANNs it is represented by a matrix of neural weights.

LCSs’ knowledge representation is closer to that of

a knowledge base, as in LCSs knowledge is mod-

elled as IF-THEN rules (with an assigned quality).

Thus, we will ﬁrst evaluate a transformation between

knowledge bases and LCSs before trying the same

with ANNs.

5 CONCLUSIONS

This work illustrated the challenges faced during

commissioning of machinery and detailed how auto-

matic parameter optimisation can decrease both time

needed for commissioning as well as down-times dur-

ing the remainder of a machine’s lifetime. We intro-

duced FDM printing as a case study, outlining param-

eters that can be adjusted to increase its robustness

against environmental factors. To achieve automatic

parameter optimisation we outlined an approach to

develop a learning system that could also facilitate

on-boarding of and knowledge transfer between com-

missioners. A concrete next step will be the devel-

opment of a pipeline to reproducibly conduct exper-

iments with FDM printers to quantify the effect of

different parameter conﬁgurations. In future research

we will strive to realise the envisioned goal by imple-

menting and evaluating the described approach on the

dataset resulting from these experiments.

ACKNOWLEDGEMENTS

This work is supported by the German Federal Min-

istry for Economic Affairs and Energy (BMWi).

REFERENCES

Abel, M. (2017). Automatisierte Inbetriebnahme von rekon-

ﬁgurierbaren Bearbeitungsmaschinen mit serviceori-

entierten Paradigmen. Stuttgart : Fraunhofer Verlag.

Alafaghani, A., Qattawi, A., Alrawi, B., and Guzman, A.

(2017). Experimental Optimization of Fused Deposi-

tion Modelling Processing Parameters: A Design-for-

Manufacturing Approach. Procedia Manufacturing,

10:791–803.

Butz, M. and Wilson, S. W. (2002). An Algorithmic De-

scription of XCS. In Soft Comput. 6, pages 144–153.

Cohn, D., Ladner, R., and Waibel, A. (1994). Improv-

ing generalization with active learning. In Machine

Learning, pages 201–221.

Das, R., Dhuliawala, S., Zaheer, M., Vilnis, L., Durugkar,

I., Krishnamurthy, A., Smola, A., and McCallum, A.

(2017). Go for a walk and arrive at the answer: Rea-

soning over paths in knowledge bases using reinforce-

ment learning. arXiv preprint arXiv:1711.05851.

Gibson, I., Rosen, D., and Stucker, B. (2015). Ad-

ditive Manufacturing Technologies - 3D Printing,

Rapid Prototyping, and Direct Digital Manufacturing.

Springer-Verlag.

Hoffmann, P., Schumann, R., Maksoud, T. M., and Premier,

G. C. (2010). Virtual Commissioning Of Manufac-

turing Systems A Review And New Approaches For

Simpliﬁcation. In ECMS, pages 175–181.

Holland, J. H. (1976). Adaptation. In Progress in Theo-

retical Biology, volume 4, page 263–293. Academic

Press, New York.

Koren, Y. (2006). General RMS Characteristics. Compar-

ison with Dedicated and Flexible Systems. Recon-

ﬁgurable Manufacturing Systems and Transformable

Factories.

Kotsiantis, S. B. (2013). Decision trees: a recent overview.

Artiﬁcial Intelligence Review, 39(4):261–283.

Kozior, T. and Kundera, C. (2017). Evaluation of the In-

ﬂuence of Parameters of FDM Technology on the Se-

lected Mechanical Properties of Models. Procedia En-

gineering, 192:463–468.

Lee, C. G. and Park, S. C. (2014). Survey on the virtual

commissioning of manufacturing systems. Journal

of Computational Design and Engineering, 1(3):213–

222.

Lehmann, D. and Magidor, M. (1992). What does a condi-

tional knowledge base entail? Artiﬁcial intelligence,

55(1):1–60.

Liao, S.-H. (2005). Expert system methodologies and

applications—a decade review from 1995 to 2004. Ex-

pert Systems with Applications, 28(1):93–103.

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

412

Lin, Y., Liu, Z., Luan, H., Sun, M., Rao, S., and Liu, S.

(2015). Modeling Relation Paths for Representation

Learning of Knowledge Bases.

Ma, Y., Peng, H., and Cambria, E. (2018). Targeted aspect-

based sentiment analysis via embedding common-

sense knowledge into an attentive LSTM. In Thirty-

Second AAAI Conference on Artiﬁcial Intelligence.

Mohamed, O. A., Masood, S. H., and Bhowmik, J. L.

(2015). Optimization of fused deposition model-

ing process parameters: a review of current research

and future prospects. Advances in Manufacturing,

3(1):42–53.

Nguyen, D. Q., Nguyen, T. D., Nguyen, D. Q., and Phung,

D. (2018). A Novel Embedding Model for Knowl-

edge Base Completion Based on Convolutional Neu-

ral Network. In Proceedings of the 2018 Conference

of the North American Chapter of the Association for

Computational Linguistics: Human Language Tech-

nologies, volume 2, pages 327–333.

Nickel, M., Murphy, K., Tresp, V., and Gabrilovich, E.

(2016). A review of relational machine learning for

knowledge graphs: From Multi-Relational Link Pre-

diction to Automated Knowledge Graph Construction.

Proceedings of the IEEE, 104(1):11–33.

Socher, R., Chen, D., Manning, C. D., and Ng, A. Y. (2013).

Reasoning With Neural Tensor Networks for Knowl-

edge Base Completion. In NIPS.

Stein, A., Maier, R., and H

ahner, J. (2017a). Toward Cu-

rious Learning Classiﬁer Systems: Combining XCS

with Active Learning Concepts. In Proceedings of the

Genetic and Evolutionary Computation Conference

Companion, GECCO ’17, pages 1349–1356, New

York, NY, USA. ACM.

Stein, A., Rauh, D., Tomforde, S., and H

ahner, J. (2016).

Architecture of Computing Systems – ARCS 2016:

29th International Conference, Nuremberg, Germany,

April 4-7, 2016, Proceedings, chapter Augmenting the

Algorithmic Structure of XCS by Means of Interpola-

tion, pages 348–360. Springer International Publish-

ing, Cham.

Stein, A., Rauh, D., Tomforde, S., and H

ahner, J. (2017b).

Interpolation in the eXtended Classiﬁer System: An

Architectural Perspective. Journal of Systems Archi-

tecture, pages 79–94.

Stein, A., Tomforde, S., Diaconescu, A., H

ahner, J., and

uller-Schloer, C. (2018). A Concept for Proac-

tive Knowledge Construction in Self-Learning Au-

tonomous Systems. In 2018 IEEE 3rd International

Workshops on Foundations and Applications of Self*

Systems (FAS*W), pages 204–213.

Sun, Z., Yang, J., Zhang, J., Bozzon, A., Huang, L.-K., and

Xu, C. (2018). Recurrent knowledge graph embed-

ding for effective recommendation. In Proceedings of

the 12th ACM Conference on Recommender Systems,

pages 297–305. ACM.

Sutton, R. S., McAllester, D. A., Singh, S. P., and Man-

sour, Y. (2000). Policy Gradient Methods for Rein-

forcement Learning with Function Approximation. In

Solla, S. A., Leen, T. K., and M

uller, K., editors, Ad-

vances in Neural Information Processing Systems 12,

pages 1057–1063. MIT Press.

Tandon, N., Varde, A. S., and de Melo, G. (2018). Com-

monsense knowledge in machine intelligence. ACM

SIGMOD Record, 46(4):49–52.

Urbanowicz, R. J., Granizo-Mackenzie, D., and Moore,

J. H. (2012). Using Expert Knowledge to Guide

Covering and Mutation in a Michigan Style Learn-

ing Classiﬁer System to Detect Epistasis and Hetero-

geneity, pages 266–275. Springer Berlin Heidelberg,

Berlin, Heidelberg.

VDW (1997). Abteilungs

ubergreifende Projektierung kom-

plexer Maschinen und Anlagen. VDW-Bericht.

Wang, H., Zhang, F., Xie, X., and Guo, M. (2018). Dkn:

Deep knowledge-aware network for news recommen-

dation. In Proceedings of the 2018 World Wide Web

Conference on World Wide Web, pages 1835–1844.

International World Wide Web Conferences Steering

Committee.

Wang, J., Wang, Z., Zhang, D., and Yan, J. (2017). Combin-

ing Knowledge with Deep Convolutional Neural Net-

works for Short Text Classiﬁcation. In IJCAI, pages

2915–2921.

Wilson, S. W. (1995). Classiﬁer Fitness Based on Accuracy.

In Evol. Comput. 3, page 149–175.

Wilson, S. W. (2000). Get Real! XCS with Continuous-

Valued Inputs. In Learning Classiﬁer Systems, page

209–219. Springer Berlin Heidelberg,.

Wohlers, T. and Gornet, T. (2014). History of additive man-

ufacturing. Wohlers report, 24(2014):118.

Yang, B. and Mitchell, T. (2019). Leveraging knowl-

edge bases in LSTMs for improving machine reading.

arXiv preprint arXiv:1902.09091.

Towards Automated Parameter Optimisation of Machinery by Persisting Expert Knowledge

413