Professional Documents
Culture Documents
1 s2.0 S2212827122002591 Main
1 s2.0 S2212827122002591 Main
com
ScienceDirect
Procedia CIRP 107 (2022) 264–269
www.elsevier.com/locate/procedia
Abstract
An accurate forecast of the final manufacturing costs in the early product development phase is a key element, especially in mechanical engineer-
ing. An accurate and reliable prediction allows, on the one hand, estimating the economic feasibility and, on the other hand, deriving necessary cost
reductions. It is essential to evaluate and monitor compliance with the cost target initially and throughout the entire development process. Most
cost estimation methods are primarily based on simple heuristic and statistical approaches that provide only limited accurate predictions in the
early development phase. Sufficiently precise cost estimations, especially for customer-specific developments, are only possible in the late phase
of the development process using finely parameterized models and cost rates at the component level. The supporting use of Machine Learning
(ML) has not yet played a major role in cost estimation in mechanical engineering. Therefore, this article shows the general applicability of ML
for predicting manufacturing costs in mechanical engineering and that even in the early phase of product development, precise predictions with
little known information about the final product are possible. In addition, we present a generic feature set for ML-based cost prediction regarding
machinery. Furthermore, the quality of different created ML models was evaluated based on real data of a mechanical engineering use case, and
high accuracy in predicting manufacturing costs was demonstrated.
© 2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the International Programme committee of the 55th CIRP Conference on Manufacturing Systems
Keywords:
Machine Learning; Artificial Intelligence; Cost Prediction; Mechanical Engineering; Product Development
This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
Christoph Hennebold et al. / Procedia CIRP 107 (2022) 264–269 265
2. Background
This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
266 Christoph Hennebold et al. / Procedia CIRP 107 (2022) 264–269
for schedule success, 80 % [35]. Ambrule and Bhirud (2017) fore, machine type independent and early known features have
used an NN model for cost prediction of reinforced concrete to be identified, suitable AI techniques have to be determined
buildings, whereas results show 93 % average accuracy with a and small training data sets in SMEs have to be considered.
mean squared error (MSE) of 0.038 [1]. Cao et al. (2018) de-
veloped a multi-layer ensemble learning model for predicting
3. Machine learning based cost prediction
unit price bids for highway rehabilitation projects [6]. The en-
semble learning model shows a mean absolute percentage er-
ror (MAPE) of 7.56. Kim (2013) incorporated Analytical Hi- As shown in Chapter 2.2 a multitude of classical and AI-
erarchy Process and case-based reasoning (CBR). They built a based approaches exist for estimating the manufacturing or con-
reliable cost estimation model for highway projects in South struction costs during the design stage. As already reviewed, the
Korea, resulting in a mean average error(MAE) of 2.76 [16]. use of AI in manufacturing cost prediction has proven feasible
Leśniak and Zima (2018) also used a CBR approach to estimate and advantageous. However, these approaches are not applica-
costs for sports field construction using 16 predictors based on ble in the early development phase in mechanical engineering
143 construction projects. The model shows a final total er- SMEs. There are major differences in features, some of which
ror of 14 % [17]. Other researchers applied regression mod- are very different from those available in mechanical engineer-
els to estimate the cost of road construction [24] or residential ing. These approaches already require a sound knowledge of
construction projects [32] achieving high prediction accuracy. the final product design, which usually is not available in the
Arabzadeh et al. (2018) developed NN, regression and hybrid early development phase. Furthermore, there isn’t a consensus
models for spherical storage tank cost estimation and showed on which AI techniques are best for cost prediction, and NNs
the best result for hybrid NNs [3]. require huge data sets that won’t be available in mechanical en-
There are some works regarding AI techniques in cost esti- gineering SMEs.
mation despite the construction business. For instance, Orlando To address these challenges, a concept of a manufacturing
et al. (2008) successfully developed an NN model for early cost cost prediction is proposed in this chapter, which is applica-
estimation of shell and tube heat exchangers [10]. Loyer et al. ble in the early development phase in mechanical engineering.
(2016) evaluated different statistical models for cost estimation This concept covers aspects, such as a generic feature set and
of jet engines and showed that gradient boosted trees and sup- the creation of a suitable data set. Furthermore, different ML al-
port vector regression are up to two times more efficient than the gorithms are considered and evaluated concerning applicability
ones typically encountered in the literature like Multiple Linear and prediction accuracy.
Regression and NNs [20]. Soyong (2020) applied a deep learn- The creation of a sufficient data set is the first part of the
ing model to estimate the manufacturing cost of mechanical concept. In this regard, selecting suitable features is very im-
components based on 3D models [37]. Due to the monolithic portant here, since they need to be available in the early devel-
and simple character of the components analyzed, the adapt- opment phase. Therefore, a generic feature set is proposed be-
ability of this approach to a complete machine is limited. cause different areas of mechanical engineering may have very
domain-specific features that cannot be applied to all areas of
2.3. State of the art analysis mechanical engineering as a whole. Table 1 shows an overview
of all features used in this proposed approach. On the one hand,
Cost estimation methods, differentiated into quantitative and features were derived from existing cost estimation methods ap-
qualitative methods, are characterized by different suitability plicable in the early development phase, such as the weight and
for specific product development phases. While quantitative dimensions of a machine which are derived from the kilo cost
methods tend to provide precise estimations in the middle and method. On the other hand, features based on early available
late development phases, qualitative methods are already appli- technical data like energy consumption and power supply were
cable at an early development stage. However, qualitative meth- selected.
ods are associated with enormous uncertainty [26, 11, 12]. But The second part of the concept comprises the selection of
especially in the early development phase, reliable cost estima- suitable prediction algorithms. In this regard, different algo-
tion is crucial for SMEs in mechanical engineering, although rithms are considered capable of dealing with small data sets.
detailed knowledge of the final product is not available. AI tech-
niques for cost estimation are established in the scientific liter-
Table 1. Used Features
ature, and preliminary work has shown that AI-based cost es-
timation methods can provide accurate results. However, these Feature Unit of measure Origin
approaches also have shortcomings regarding the applicability {Length, Width, Height} mm Kilo Cost Method
in the early phase of product development in mechanical en- Weight kg Kilo Cost Method
gineering. The bespoken approaches require sufficient knowl- Machine Type {A, B, C, D} technical data
edge about the final “product” like construction plans or de- Power Supply Volt (V) technical data
Energy Consumption kw technical data
tailed 3D models [37]. Consequently, neither the current mod- Performance pcs/h cost function
els nor the specific feature engineering can be directly applied Level of Automation {manual, semi, automatic} software
to the use case of technical complex machinery due to the lack Manufacturing Cost € software
of product-specific data in the early development phase. There-
This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
Christoph Hennebold et al. / Procedia CIRP 107 (2022) 264–269 267
Therefore, the linear regression (LR), decision tree (DT), ran- Table 2. Cost prediction results.
dom forest (RF), K-nearest neighbor (KNN), and support vec-
Algorithm MAE RMSE
tor regression (SVR) are considered. The prediction accuracy of
those algorithms in the context of the mechanical engineering KNN €672 ±€451 €703 ±€449
use case was tested and compared during validation (cf. Sec- LR €994 ±€625 €1178 ±€801
tion 4.2). The respective results were evaluated by means of the RF €2,675 ±€1,990 €3,073 ±€2,412
SVR €2,711 ±€2,031 €3,042 ±€2,351
MAE and root rean squared error (RMSE) metrics in order to DT €2,742 ±€2,273 €3,232 ±€2,952
be able to compare their applicability and forecasting accuracy.
4. Validation the mean deviation and, second, the effect of possible outliers
on the cost prediction. The calculated error metrics for each
In this chapter, the validation of the proposed ML-based ap- method are shown in Table 4.2.
proach for manufacturing cost prediction presented in Chap- The application of the KNN algorithm results in an MAE of
ter 3 is described. The approach was validated by means of a €672, showing the best prediction accuracy of all compared al-
use case of a mechanical engineering company. In this regard, gorithms, which is due to the property of KNN to average the
a data set was created on the basis of different packaging ma- results of the k-nearest neighbors for decision-making. Even if
chine types from cardboard processing and represent cardboard the predicted manufacturing costs increase, the averaged value
sealing, cardboard erecting, shrinking, and strapping machines leads to a more accurate result. A similar result occurred with
which are respectively abbreviated from A to D in Table 1. First, the LR algorithm, which is slightly higher with a MAE of €994.
the created dataset, correlations in the data and peculiarities in It is noticeable that the DT, RF and SVR algorithms predicted
the used features are described. Based on the created data set, manufacturing costs in a similar range with an MAEs between
the regression algorithms mentioned in Chapter 3 are tested for 2,600 - €2,750 and are also suitable to determine the manu-
their applicability in manufacturing cost prediction. Moreover, facturing costs. It is worth noting that the results of KNN and
the cost prediction results are evaluated using the MAE and LR differ greatly from those of RF, SVR, and DT, forming
RMSE metric two different clusters of algorithms. In this context, observing
the RMSE metric shows the calculated error (deviation) of the
4.1. Dataset model predictions from the target value. It provides additional
information about possible outliers in the predictions, which
The data set consists of 81 data points comprising the negatively affect the computed value. As can also be seen in Ta-
generic features mentioned in Table 1 as well as the manu- ble 4.2, across all algorithms, the calculated RMSE values lie
facturing costs of the respective machines as target. The data close to the manufacturing costs computed in MAE. The RMSE
set was compiled manually in the course of this work and is value of all algorithms is about €200 to €600 higher, which can
based on information from product data sheets and the website be explained with possible outliers in the calculations as previ-
content of the use case company. Occasionally, features of data ously described. The application of the existing test data for the
points were not completely available. To fill in the missing fea- prediction of manufacturing costs has shown that ML can be
tures based on existing data points, an iterative imputer with applied for cost prediction in the field of mechanical engineer-
Bayesian Ridge Regression estimator was utilized. The maxi- ing. However, the very small size and the unbalanced nature
mum number of imputation rounds was set to 10 and missing of the data set is a problem in calculating manufacturing costs.
values were initialized with the computed mean. The manu- The very small amount of training data does not allow the algo-
facturing costs stored in this data set are in a price range be- rithms to reasonably learn the transitions between the price seg-
tween approx. 500–€28,000. The correlation matrix of the cre- ments of different machines. This effect is exacerbated by the
ated dataset in Figure 4.1 shows the correlations of the used unbalanced nature of the data set. It contains machines from a
data. The high correlation between the dimensions and the ma- wide range of price segments, although not every price segment
chine’s weight is not surprising. In addition, depending on the is represented by the same amount of machines. However, in-
size and weight, its power supply also changes, reflected in a creasing the size of the data set with additional data seems very
relatively high correlation. It is also noticeable that the degree worthwhile to improve the predictive accuracy of the methods
of automation of a machine affects its size and therefore also further. Moreover, the enlargement of the database also enables
its weight, whereby the size and thus also the weight increases utilizing deep learning methods, which were not yet considered.
with an increasing degree of automation.
To validate the applicability of the different regression algo- This article presented a novel and ML-based approach to
rithms for manufacturing cost prediction and to evaluate their predict manufacturing costs of new products during the early
prediction accuracy, they are examined by 10-fold cross val- phase of product development of mechanical engineering SMEs
idation on the metrics MAE and RMSE to investigate, first, comprising a generic feature engineering. Due to that, the ap-
This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
268 Christoph Hennebold et al. / Procedia CIRP 107 (2022) 264–269
This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
Christoph Hennebold et al. / Procedia CIRP 107 (2022) 264–269 269
[6] Cao, Y., Ashuri, B., Baek, M., 2018. Prediction of unit price bids of resur- Sustainable Real Estate and Construction Economics 1, 182–199.
facing highway projects through ensemble machine learning. Journal of [25] Ozbayrak, M., Akgun, M., Turker, A.K., 2004. Activity-based cost estima-
Computing in Civil Engineering 32, 04018043. tion in a push/pull advanced manufacturing system. International Journal of
[7] Christensen, C.M., 2013. The innovator’s dilemma: when new technologies Production Economics 87, 49–65. URL: https://EconPapers.repec.
cause great firms to fail. Harvard Business Review Press. org/RePEc:eee:proeco:v:87:y:2004:i:1:p:49-65.
[8] Chwastyk, P., Kołosowski, M., 2014. Estimating the cost of the new [26] PMI, A., 2008. December. Guide to the project management body of
product in development process. Procedia Engineering 69, 351–360. knowledge: Pmbok 2000. project management institute.
URL: https://www.sciencedirect.com/science/article/pii/ [27] Schuh, G., Riesener, M., Dölle, C., 2016. Concept for development project
S1877705814002458, doi:https://doi.org/10.1016/j.proeng. management by aid of predictive analytics, in: 2016 Portland International
2014.02.243. 24th DAAAM International Symposium on Intelligent Conference on Management of Engineering and Technology (PICMET),
Manufacturing and Automation, 2013. pp. 2040–2047. doi:10.1109/PICMET.2016.7806640.
[9] Coenenberg, A.G., Fischer, T.M., Günther, T., 2016. Kostenrechnung und [28] Sequeira, S., Lopes, E., 2015. Simple method proposal for cost
kostenanalyse. Schäffer-Poeschel. estimation from work breakdown structure. Procedia Computer
[10] Duran, O., Rodriguez, N., Consalter, L.A., 2009. Neural networks for cost Science 64, 537–544. URL: https://www.sciencedirect.
estimation of shell and tube heat exchangers. Expert Systems with Appli- com/science/article/pii/S1877050915026940, doi:https:
cations 36, 7435–7440. //doi.org/10.1016/j.procs.2015.08.559. conference on EN-
[11] Hegazy, T., Ayed, A., 1998. Neural network model for parametric cost esti- TERprise Information Systems/International Conference on Project
mation of highway projects. Journal of Construction Engineering and Man- MANagement/Conference on Health and Social Care Information Sys-
agement 124, 210–218. doi:10.1061/(ASCE)0733-9364(1998)124: tems and Technologies, CENTERIS/ProjMAN / HCist 2015 October 7-9,
3(210). 2015.
[12] International, A., 2004. Recommended practices. https://web.aacei. [29] Shehab, E.M., Abdalla, H.S., 2002. A design to cost system for innova-
org/resources/recommended-practices. tive product development. Proceedings of the Institution of Mechanical
[13] John Wiley Sons, L., 2011. The System Development Pro- Engineers, Part B: Journal of Engineering Manufacture 216, 999–1019.
cess. URL: https://onlinelibrary.wiley.com/doi/abs/ [30] Siadat, A., Dantan, J.Y., Mauchand, M., Martin, P., 2007. A cost estima-
10.1002/9781118001028.ch4, doi:https://doi.org/10.1002/ tion system for manufacturing product using ontology and exepert system.
9781118001028.ch4. IFAC Proceedings Volumes 40, 37–42.
[14] Jung, J.Y., 2002. Manufacturing cost estimation for machined parts based [31] Singal, P., Kumari, A.C., Sharma, P., 2020. Estimation of soft-
on manufacturing features. Journal of intelligent manufacturing 13, 227– ware development effort: A differential evolution approach. Pro-
238. cedia Computer Science 167, 2643–2652. URL: https://www.
[15] Kim, G.H., Shin, J.M., Kim, S., Shin, Y., 2013. Comparison of school sciencedirect.com/science/article/pii/S1877050920308097,
building construction costs estimation methods using regression analysis, doi:https://doi.org/10.1016/j.procs.2020.03.343. interna-
neural network, and support vector machine . tional Conference on Computational Intelligence and Data Science.
[16] Kim, S., 2013. Hybrid forecasting system based on case-based reasoning [32] Stoy, C., Pollalis, S., Dursun, O., 2012. A concept for developing construc-
and analytic hierarchy process for cost estimation. Journal of Civil Engi- tion element cost models for german residential building projects. Interna-
neering and Management 19, 86–96. tional Journal of Project Organisation and Management 4, 38–53.
[17] Leśniak, A., Zima, K., 2018. Cost calculation of construction projects in- [33] Sullivan, F., 2021. Increasing competitive intensity to push major suppliers
cluding sustainability factors using the case based reasoning (cbr) method. out of business in the north american brakes aftermarket.
Sustainability 10, 1608. [34] Wang, Y.j., Yang, J.x., Shi, J.l., 2011. Study on activity-based cost esti-
[18] Lin, C.K., Shaw, H.J., 2017. Feature-based estimation of preliminary costs mation of steel enterprise, in: 2011 IEEE 18th International Conference
in shipbuilding. Ocean Engineering 144, 305–319. URL: https://www. on Industrial Engineering and Engineering Management, pp. 968–972.
sciencedirect.com/science/article/pii/S002980181630542X, doi:10.1109/ICIEEM.2011.6035317.
doi:https://doi.org/10.1016/j.oceaneng.2016.11.040. [35] Wang, Y.R., Yu, C.Y., Chan, H.H., 2012. Predicting construction cost
[19] Lin, Z.C., Chang, D.Y., 2002. Cost-tolerance analysis model based and schedule success using artificial neural networks ensemble and sup-
on a neural networks method. International Journal of Produc- port vector machines classification models. International Journal of Project
tion Research 40, 1429–1452. URL: https://doi.org/10. Management 30, 470–478.
1080/00207540110116282, doi:10.1080/00207540110116282, [36] Yadav, N., Gupta, N., Aggarwal, M., Yadav, A., 2019. Comparison of
arXiv:https://doi.org/10.1080/00207540110116282. cosysmo model with different software cost estimation techniques. 2019
[20] Loyer, J.L., Henriques, E., Fontul, M., Wiseall, S., 2016. Comparison of International Conference on Issues and Challenges in Intelligent Comput-
machine learning methods applied to the estimation of manufacturing cost ing Techniques (ICICT) 1, 1–5.
of jet engine components. International Journal of Production Economics [37] Yoo, S., Kang, N., 2021. Explainable artificial intelligence for manufactur-
178, 109–119. ing cost estimation and machining feature visualization. Expert Systems
[21] Mahadik, A., Masel, D., 2018. Implementation of additive man- with Applications , 115430.
ufacturing cost estimation tool (amcet) using break-down ap-
proach. Procedia Manufacturing 17, 70–77. URL: https://www.
sciencedirect.com/science/article/pii/S2351978918311302,
doi:https://doi.org/10.1016/j.promfg.2018.10.014. 28th
International Conference on Flexible Automation and Intelligent Man-
ufacturing (FAIM2018), June 11-14, 2018, Columbus, OH, USAGlobal
Integration of Intelligent Manufacturing and Smart Industry for Good of
Humanity.
[22] Naranje, V., Kumar, S., Hussein, H., 2014. A knowledge based system for
cost estimation of deep drawn parts. Procedia Engineering 97, 2313–2322.
[23] Niazi, A., Dai, J.S., Balabani, S., Seneviratne, L., 2005. Product Cost
Estimation: Technique Classification and Methodology Review. Journal
of Manufacturing Science and Engineering 128, 563–575. URL: https:
//doi.org/10.1115/1.2137750, doi:10.1115/1.2137750.
[24] Ogungbile, A.J., Oke, A.E., Rasak, K., 2018. Developing cost model for
preliminary estimate of road projects in nigeria. International Journal of
This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.