Available online at www.sciencedirect.com [626382]

Available online at www.sciencedirect.com
2212-8271 © 2016 The Authors. Published by Elsevier B.V . This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer
-review under responsibility of the scientific committee of the 5th CIRP Global Web Conference Research and Innovation for Future Production
doi: 10.1016/j.procir.2016.08.014 Procedia CIRP 55 ( 2016 ) 212 – 217 ScienceDirect
5th CIRP Global Web Conference Research and Innovation for Future Production
Development of a knowledge-based predictive model to estimate
the welding process time in single part production systems
Farhang Akhavei*, Arameh Khallaghi, Friedrich Bleicher
Institute for Production Engineering and Laser Technology, Vienna University of Technology; Vienna
* Corresponding author. Tel.: +43 1 58801 31117; fax: +43 1 58801 31199; E-mail address: [anonimizat]
Abstract
The lack of repetition effect in the single part production in metal industry limits an exact determination of the necessary pr ocess parameters (e.g.
welding time) for production planning and furthermore restricts the application of production planning systems. This work discu sses a
methodology to improve the prediction accuracy of welding time in single part production systems. In order to determine the act ual process
parameters of simple welding process characteristic indicators are identified. These indicators are stored as process features and used during the
process planning phase. To achieve this target, the configuration and integration of a Product Data Management (PDM) and a Business
Intelligence System is necessary. In the case of simple welding processes these indicators are calculated through mathematical for m ulas, which
are developed on the basis of welding parameters. But this methodology can’t be used for complex welding processes with many in fluence
factors. For this kind of processes other prediction models are developed on the basis of analytics methodology. After validati on of these models,
they are integrated in a data warehouse system and work automatically within a knowledge-based circuit. The accuracy of the ind icators
continuously improves through newly acquired data (learning effect). This methodology supports single part producers to improve their
production planning quality in metal industry.
© 2017 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the scientific committee of the 5th CIRP Global Web Conference Research and Innovation for Future
Production.
Keywords: knowledge-based system, data warehouse system, single part and small series production, welding process, predictive analytics
methodolog
y

1. Introduction
Th
e reliability of production planning plays a critical role for
the effectiveness and efficiency of modern production systems.
The Accuracy of the determination of production planning parameters like process and setup time is also a strong
requirement for the reliability of production planning. The
differences between planning parameters and the real
production parameters reduce the accuracy and reliability of
production planning. Schuh et al. have shown that on average
the deviation of the planning parameters may occur on 25%in only three days after system validation [1]. The lack of high
repetition effect in the single part production in metal industry
limits an exact determination of the necessary process parameters like process time in production planning. This problem is especially noticeable in welding processes. In single
part production, often have the di fferent parts in a work station
different dimensions. Therefor the estimation of process time
is a big challenge in production planning. The actual state of art
includes only few methods for the prediction of welding
process time, which are based on welding technologies. For
example, Masmoudi et all. suggest a method to estimate the
welding cost and time based on feature concept [2]. Heimbokel
has also suggested the similar approach based on technological aspects of welding to determin e the welding process time [3].
The challenge for the application of these technological approaches, is a very high level of complexity and a low level of flexibilities. Therefore, these approaches are normally
suitable for automated welding processes. To implement these
methods, is also the determination of many technological parameters necessary, which is normally a big practical
© 2016 The Authors. Published by Elsevier B.V . This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer
-review under responsibility of the scientifi c committee of the 5th CIRP Global Web Conference Research and Innovation for Future Production

213 Farhang Akhavei et al. / Procedia CIRP 55 ( 2016 ) 212 – 217
challenge in a real production process. Furthermore, in these
approaches, the logistical, ergonomic, organizational and other
process aspects like parts preparati on or plate cleaning can’t be
considered. Therefore, the application of these approaches in single part production systems is limited. Due to the high
impact of these mentioned influence aspects in the manual
work stations compared to auto mated processes, this challenge
is also bigger in manual welding process and therefore the
determination of process time only based on the application of
technological welding parameters like welding performance is
very inaccurate. In this work we introduce the predictive
modelling method as a possible solution. Through our
approach, an output parameter, in our case process time, is
estimated based on mathematical relation and correlation with different entry parameters. These entry parameters are in this
case, design and construction parameters, which are
determined after the product design phase. The historical operating data are used as basis to develop the predictive
model. In this methodology the simple regression models are
used as a predictive tool. In actual state of the art, some works can be found, which present the deferent application of
predictive analytics in series production specially through the
predictive modelling on basis of big data with the algorithms
like machine learning. But due to the limited production orders
in single part production systems compared to series production, the application of predictive models based on big
volume of data is in this sector very limited. Therefore, the focus of the methodology in this work lies on the development
of simple predictive models for a practical application in single
part production to optimize the accuracy of welding process production planning. The application of characteristic indicators to increase prediction accuracy enables the
optimization of predictive modelling and is deemed a novel
approach. The major advantage of this methodology is its high flexibility. This methodology can be also carried out only on
the basis of historical data and it is not necessary to apply
another process knowledge. Furthermore, all process parameters like technical, logistical, organizational and
ergonomic are automatically cons idered in this solution.
Therefore, compared to technical methods to predict the
welding process time, more accuracy is expected from the
presented methodology. After validation, the method is
implemented in the software landscape of the company to
automate the optimization of the knowledge-based model
which in combination with the beforementioned factors is
another scientific novelty of the approach.

2. Requirements

The current work describes a method to predict the welding
pr
ocess time in single part pr oduction systems. Due to the
process variety, the sort of requirements must be fulfilled in
order to generalize the methodology and to make it compatible
with our system demands. At the first step, the standardization is the main requirement to apply this methodology. Through
this:
1- The workers, process, work tools, welding
technology, logistic flow and work fl
ow in the same
work station must be consistent and should have the same quality and quantity as the pervious case.
2- In case of different types of welding technologies,
welding
processes must be classified. 3- The new components and assemblies must be in the s
ame product family and have similar or comparable
forms (not dimension). Often the different product in different product family require the different
production process, manufacturing technology and process flow. Therefore, it is necessary to classify the
components and assemblies based on their
characteristics.
4- The same components and assemblies must be
produce
d at the same work station.

To implement this methodology , t he required information
systems like ERP, BI and PDM should be also available in IT landscape of company.
3. State of the art
3.
1. Knowledge-based production planning
A
knowledge-based system is a program that uses knowledge
to solve complex problems with attempting to represent
knowledge explicitly via tools. Th is system can be used in
many different fields. It is al so one of the major members in
Artificial Intelligence group [4]. The k nowledge-based
decision support systems are a new generation of DSS, which
work on the basis of historical data. These systems can also support the production planning and scheduling to increase the
planning reliability. 1994 Hinkelma nn et al. have presented the
knowledge-based product and production planning as a significant evolution in production, which can optimize the production in a high level [5]. Schuh et al. have shown that the reliability of the production planning can be increased by the intelligent application of production data in a knowledge-based
system [6]. They have introdu ced in their work a knowledge-
based system as an important component of an autonomous and self-organized production system . They have also applied the
optimization concept from Horn et al. to improve the scheduling and sequencing by automatic model generation (AMG) [7]. AMG provides a continuous adjustment of the simulation and planning models with the current system behaviour to increase the accuracy of the production planning [8]. Bubeník et al. have also presented an approach for knowledge-based production pl anning by applying the
knowledge indicators. In their model, the current design and
production parameters are save d in a DWH (Data warehouse)
system and transformed into re levant planning indicators.
These indicators are used to improve the accuracy and reliability of the production planning constantly, through predict the waste rate of production [9].
3.2. Determining the welding process time in single part
production
In
the current state of the art, some works can be found, which
introduce different approaches to predict the process time. For
example, Müller has suggested the following three methods to estimate the process time: experiential knowledge (standard or estimated time), historical data and mathematical functions [10]. On this base, he has developed a hybrid method to estimate the process time in the production. He has

214 Farhang Akhavei et al. / Procedia CIRP 55 ( 2016 ) 212 – 217
demonstrated the experience and knowledge method for new
technologies, historical data for established processes and
mathematical functions for interdisciplinary production processes. Faisst et al. have also presented an approach based on a mathematical forecasting system to predict the process time [11]. This prediction system constitutes a learning effect through a larger data base, which improves the accuracy of the results continuously. In the current work, their approach is respected as a basis to impr ove the production planning in
single part and small batch production. Luehe has also introduced the similar method to estimate process cost by applying mathematical methods [12]. He has used a modular system to standardize assemblies and parts and has applied stochastic functions based on the acquired process data to determine the process cost. The calculated values are deposited behind the respective modules (as features) and this acquired knowledge is applied in the ne xt project planning. The model
of Seung-Jun et al. can also be viewed as a very interesting and relevant approach for determining the production parameters in the complex production systems [13]. However, the model of Seung-Jun et al. was developed to predict the energy consumption in production systems, but the model, the methodology and the approach can be also used to predict the planning relevant production pa rameters. They have applied
the analytics method based on big data and machine learning to develop the forecast model for energy consumption in manufacturing. In their model, the correlations of the input parameters like material, machine tools etc. with a determined unit of energy (output parameters) are analysed and the predictive model has been developed through the neural network techniques. The model is developed based on the large volume of data sets (10,000 records).

3.3. Predictive analytics
Predictive anal
ytics is described as the use of a variety of tools
(predictive modelling, machine l earning, data mining) to make
predictions about future events based on historical events [14]. Predictive analytics is carried out on the basis of predictive
modelling. This modelling is also described as a process of developing the mathematical tool or model that generates an accurate prediction [15]. Predictive analytics is also a technology, that learns from data to predict the future behaviour of individuals. Its application includes different aspects like:
price prediction, risk assessment, etc. [16].
4. Methodology
4.
1. Describing the methodology
To
develop a methodology for the described problem statement
in this work, the approaches of Faisst et al. and Luehe are respected and in consideration of Seung-Jun et al. statement, the predictive modelling is applied as analytics tool. But it should be noted, because of descri bed situation in single part
production, in this methodology has been tried to reduce the
model complexity through develop the simple predictive models. In the next step and to increasing the model accuracy and also to simplify the mode lling, this approach has been optimized through the characteristic (specific) indicator methodology in consideration of Bubeník et al approach.
These indicators describe the rate of work and facilitate the storage of process and project knowledge in a neutral form to use it in future. Welding performance is for example a technological characteristic indicator. In the following a basic flowchart is used to illustrate the methodology. The steps of the
flow chart are expressed as follows.

4.Data analysis to select
related entry
parameters4.Data analysis to select
related entry
parameters
7.Is direct
prediction more
accurate1.Data acquisition
2.Data
classification
3. Prediction
through
characteristic
indicator 3. Direct
prediction
6.Reliability
test6.Reliability
test5.Data modelling 5.Data modelling
8.Direct
prediction 8.Prediction
through char.
indicator
9.ENDYes
No

Fig. 1. methodology flowchart
1. Collecting data in the right structure is a main requirement
to apply this methodology 2. In this step the data are classified. This classification could be based on material or welding technology, etc. which can be selected according to some other factors such as kind of product
family. 3. The data sets can be imported in two separate models to estimate the process time. The fi rst one is direct prediction. In
this case the process time as output parameter is predicted in a predictive model directly. The ot her one is prediction through
indicator, which predicts the process time indirect through specific and characteristic indicat or like weldin g speed. This
rate based indicator is determined with integration of process time as output parameter and a dimension described entry parameter like length. This indicator characterizes the new output parameter in predictive analyse. The suitable input parameter is selected based on interpretation of correlation analysis. 4,5,6. There are various analytics software for analysing and predictive modelling. The modelling techniques depend on situation and operation conditions , which could be categorized
in two different methods, with and without preselection entry parameters. To determine the model validation and reliability, the cross validation techniques is applied. In this method, 75%
of historical data sets are used for modelling (training sets) and

215 Farhang Akhavei et al. / Procedia CIRP 55 ( 2016 ) 212 – 217
25% of them are used for validation test (testing sets) [17]. The
outcome of the reliability test is a model deviation, that is the main valuation factor to compare the methods and their accuracy with each other. 7. This step comes up with the selection between direct and indirect models. In this step, the model deviations of both methods are compared and the better model is selected. 8. The selected model is applied and can be implemented in IT landscape.
4.2. Software landscape and aut omated
knowledge-based
circuit This methodology can be developed in the next step in a
knowledge-based circle to optimize the production planning continuously. To achieve it, the adaption and configuration of software landscape is required. In the first step, it is necessary that the design and construction parameters as entry parameters and the detected process time as output parameter are saved and archived in the standard structure based on determined data classes. In our approach, the DWH System is responsible for this task. After finishing th e production order in the work
station, the relevant entry parameters, which are in different positions of parts list and proce ss time as output parameters are
transferred and saved in the DWH System. Through this approach after each production order in the work station a new data set is created, which will be used to develop the predictive model. An analytics software is also responsible for the definition of a predictive calculation formula based on a predictive model. Because of the high level of complexity, the predictive model in our approach is selected and defined manually and static. Normally the information about the production process and work stations are saved and administrated in the ERP system. Therefor it would be probably suitable to configure and apply the predictive formula in the ERP system. The software-interface between BI, Analytics and ERP system suppo rts the automation of the
knowledge-based circuit. In this approach the predictive formula should be updated in defined time units and it is required to define how many data sets should be transformed in predictive model every time . The adjustment of the
predictive formula can take place automatically or through operator instruction. For the app lication of specific indicators
to predict the process time, calculation of sum welding length of assemblies is necessary. This calculation can be carried out in a PDM system and transformed in an ERP system through additional vertical position of parts list.
5. Case study
This case
describes the production of container’s cylinders in a
work station with standard pr ocedure. Cylinders are welded
here with TIG process. The process time is determined here as
sum of preparation and welding time. The control process is a separate process with the cons tant process time for all orders.
The operating data sets in this case are gathered about 3 months and includes 80 historical data sets. Each data set contains material (here S235 and MA75), diameter, length & sheet thickness of each cylinder, which are different in various projects. In this case the data sets should be classified only based on material. In the first model (direct prediction) process times are calculated direct on the basis of entry parameters and in the second model (indirect prediction), process times are calculated through a specific indicator. This indicator is identified as welding speed and describe the rate, which defines how much meters are welded in an hour. Between all tested regression models, the polynomi al model has shown the best
average R² and therefore, this model has been applied to develop the predictive model. A max. average model deviation of 10% is adopted as a goal for the predictive model.
5.1. Direct prediction
After classifica
tion, three quarters of data sets are used through
a simple regression model for analysis and modelling and the
model has been validated with the rest of the data sets (10
records). The goal of data analysis is the determination of suitable entry parameters for predictive modelling. In this case, it is adopted, that each parameter has been selected, which has
a coefficient of determination (R²) over 0,5 with process time.

Material S235: Th
e following table shows (Table 1.) the results
of data analysis:

Table 1. S235 analysis results
Fig. 2. S235 Sheet
thickness / Process time
As a result, only sheet thickness has been selected for material
S235.

Material MA75 : Th e following table (Table 2.) shows the
results of data analysis:
Table 2. MA75 analysis results

Fig. 3. MA75 Sheet thickness / Process time

As result for material MA75, at first view the sheet thickness
and diameter are suitable, but after advanced
analysing, it has been found out that there is a direct correlation
between sheet thickness and diameter (R²=0,6), therefore, to
simplify the predictive model only sheet thickness has been
selected as final entry parameter (because of bigger R²).
Entry
parameter R²
Length 0.236
Diameter 0.227
Sheet
thickness 0.879
Entry parameter R²
Length 0.231
Diameter 0.718
Sheet
thickness 0,842

216 Farhang Akhavei et al. / Procedia CIRP 55 ( 2016 ) 212 – 217
The accuracy of the model is defined by the average deviation
of predicted times from measurement times. The estimated
process times are calculated through the regression formula of
selected entry and output parameter. For example, the process
time for S235 (Fig.2.) is calcu lated with the following formula
based on sheet thickness (as X in formula [mm]):

Process time = (0,1619 × X ^ 2) + (0,2555 × X) +38,291

following table (Table 3) and figures 6 and 7 shown the average deviations
of predictive models:
Table 3. average deviation of direct prediction modelling
5.2. Prediction through characteristic indicator
In t
his method the approach and conditions are the same as in
the first method. Distinguishing is, that in this model the length
parameter and process time are transformed to a new specific output parameter, which defines the rate of welding in an hour (welding speed). The difference of this indicator with technical welding performance is it, that this indicator is specific for this
work station and is determined only based on historical data
sets. It includes also all organi sational and logistical aspects.
Due to the strong correlation between sheet thickness and process time, for simplify the model only sheet thickness has been used as entry parameter. Following figures and table shown the results of data analy sis for each material (Fig. 4, 5
and Table 4, 5):
Table 4. S235 analysis result
through indicator

Fig. 4. S235 Welding speed / Sheet thickness

Table 5. MA75 analysis results
through indicator

Fig. 5. MA75 Welding speed / Sheet thickness

After estimating of indicator, the estimated process time is
calculated through dividing of th e length by the indicator. Due
to strong correlation between diameter and sheet thickness is the length parameter, the only independent entry parameter,
that can be used to develop the characteristic indicator. Very
high concentration of data sets in only 20% of possible range
of length (73% of data sets are in range between 2 and 2.5
meter) can be explained as a reason for poor correlation of length parameter with process time.
This statistical situation
facilitates the development of predictive model, which can show in some cases (for example for MA75) an acceptable accuracy without consideration of length parameter (Table 3).
That is very surprising aspect of analysis. The following figures
(Fig. 6,7) shown the validation re sults and compare the results
of the both approaches.

Fig. 6. S235 Comparison of model reliability

Fig. 7. MA75 Comparison of model reliability
Table 6. average deviation of indirect prediction modelling

As expected, due to combination of length parameter and process tim
e in methodology, a better results have been
achieved compared to direct prediction method .
5.3. Direct prediction without preselection of entry
par
ameters
It is mentioned that there are two different approaches for
analysing and modelling after data classification, modelling
with and without preselection of en try parameters. As a part of
work, we have also tried the modelling without preselection of
entry parameters. This attempt has been carried out two times by using of generalized linear model (GLM) in Spider
software. The first one was carried out without data
Material Average deviation Deviation Min/Max
S235 12% 1% / 29%
MA75 8% 1% / 22%
Entry
parameter R²
Sheet thickness 0.915
Entry
parameter R²
Sheet thickness 0,894 Material Average deviation
(indicator) Deviation Min/Max
(indicator)
S235 9% 4% / 18%
MA75 5% 0% / 12%

217 Farhang Akhavei et al. / Procedia CIRP 55 ( 2016 ) 212 – 217
normalization and the second one w ith it. The following figures
demonstrate the results.
a b

Fig. 8. model reliability without preselection of entry parameters (a) MA75;
(b) S235
Table 7. average deviation of direct prediction model without preselected
entry parameters

The results of direct prediction model without preselection of
en
try parameters shown significantly that this approach with
GLM techniques is inaccurate compared to modelling with preselected entry parameters. Probably the number of historical data sets was not sufficient to lead us to a better results.
6. Conclusion
The
results of the case study have shown, that predictive
modelling can be applied princip ally as effective solution to
predict the welding process time in single part production. As
expected, it has also shown, that the predictive modelling based
on characteristic indicator leads to better results than direct predictive. Due to the results of case study, it can be interpreted, that the predictive modelling with manual preselected entry parameters combine with application of simple regression models and model optimization with characteristic indicators, is an effective approach in case of a small number of data sets. The application of other predictive algorithms to modelling without preselected entry para meters, can probably produce
better results as GLM, but pri ncipally the small number of data
sets is a challenge for modelling without preselected entry parameters. A very surprising aspect of this study, was the weak correlation between welding length and process time in selected manual welding work station, so that the process time can be predicted in average with 90 % accuracy (both materials) without contemplation of welding length. The reason is the concentration of 73% of data sets in only 20% of possible length range. Therefore, in this work station, there is a
low statistical importance of lengt h parameter compared to
sheet thickness. Nerveless, it has been shown, that the
integration of length in the modelling (indirect/indicator) can increase in average the model accuracy to 93%. Development of specific forecast model for each work station with high accuracy and possibility of automation the optimization circle, is exactly the big advantage of th is approach. Due to the very
high standardization of automated welding work station, the
actual technological approach based on welding performance is suitable. But the described approach can be also applied for automated welding processes and shows especially its advantage by flexible automated welding processes.
References
[1] G. Schuh, T. Potente, C. Thomas and A. Hauptvogel,
Cyber-Physical Production Management in Advances in
Production Management Systems. Sustainable Production and Service Supply Chains, State College, PA,
USA; 2013. P. 477-484
[2] F. Masmoudi, W. Hachicha and Z. Bouaziz, Development
of a welding cost estimation model based on the Feature concept in Advances in Production Engineering &
Management; 2007.
[3] J. Heimbokel, Costs in welding, „Kosten in der
Schweißtechnik“, Linde AG , Gelsenkirchen; 2014.
[4] A. Rajendra and S. Priti, Knowledge-Based Systems,
Jones & Bartlett Learning; 2010.
[5] K. Hinkelmann, M. Meyer and F. Schmalhofer,
Knowledge Base evolution for Product and production
planning, AICOM vol. 7 Nr. 2; 6 1994. P. 98-113
[6] G. Schuh, T. Potente and A. Hauptvogel, Methodology
for the evaluation of foreca st reliability of production
planning systems, in Proceedings of the 47th CIRP
Conference on Manufacturing Systems; 2014.
[7] S. Horn, G. Weigert and S. We rner, Data coupling
strategies in production environments., in 28th International Spring Semina r on Electronics Technology:
Meeting the challenges of electronics technology
progress, Piscataway, NJ; 2005. P. 278-282
[8] R. O. Fowler JW, Grand Challenges in Modeling and
Simulation of Complex Manufacturing Systems, in simulation: Transactions of The Society for Modeling and
Simulation International; 2004. P. 469-476
[9] P. Bubeník and F. Horák, Knowledge-based system to
support production planning, Tehni țki vjesnik 21; 03
2
014. P. 505-509
[10] S. Mueller, Methodology for the development and
planning-based generation and evaluation of alternative production, „Methodik für die entwicklungs- und
planungsbegleitende Generierung und Bewertung von Produktionsalternativen“, München: Herbert Utz Verlag;
2008.
[11] D. Dudic, Model for the factory lifecycle-oriented
product planning and development, „Modell für die Fabrik Life Cycle-orientier te Produktplanung und –
entwicklung“, Stuttgart: Jost-Jetter Verlag; 2010.
[12] C. Luehe, Modular cost estimates in support of
investment planning for the offer in early Basic
Engineering Phase, „Modulare Kostenschätzung als
Unterstützung der Anlagenplanung für die Angebots- und
frühe Basic Engineering Phase“, Berlin University of
Technol
ogy; 2013.
[13] S. Seung-Jun, W. Jungyub and R. Sudarsan, Predictive
analytics model for power consumption in manufacturing, in 21st CIRP Conference on Life Cycle Engineering; 2014.
[14] C. Charles Nyce, Predictive Analytics White Paper,
AICPACU; 2007.
[15] M. Kuhn and K. Johnson, Applied Predictive Modeling,
USA: Springer Science & Business Media; 2013.
[16] E. Siegel, Predictive Analy tics: The Power to Predict
Who Will Click, Buy, Lie, or Die, John Wiley & Sons,
USA, 2016.
[17] J. Starkweather, Cross Validation techniques in R: A
brief overview of some methods, packages, and
functions for assessing predi ction models, Research and
Statistical Support consu ltant, Texas, USA; 2011.

Material Average deviation Average deviation
(normalized)
S235 20% 24%
MA75 12% 22%