University of Erlangen-Nuremberg, Institute for Manufacturing Science

Cost Estimation for Large Scale Production of Sheet Metal Parts

Using Artificial Neural Networks*
Manfred Geiger; Jürgen Knoblach; Frank Backes
Abstract:
In the field of sheet metal working small and medium sized enterprises (SME) are normally supplier companies. They have
to submit valid offers for manufacturing jobs within a short time. The supply price is decisive for getting the manufacturing
job. Therefore, a method of quotation costing has been developed to increase the accuracy of cost planning for sheet metal
parts. A software system is being developed which implements the algorithms of this method starting with the analysis of
the CAD-model. The manufacturing of sheet metal parts in large scale production is mostly done with progressive dies, the
cost estimations for such tools are normally based on the plate sequence. To avoid the time and the costs of designing the
plate sequence before getting the job, a special calculation method based on Artificial Neural Networks is used for the cost
estimation of progressive dies. In connection the direct labour costs are calculated by accounting the estimated resource
consumption for tool manufacturing. Fast and precise pricing is done by the system under consideration of material costs,
workcenter costs, overhead costs and pricing strategy.

Keywords: Production Management, Cost Estimating, Sheet Metal Part

*
Introduction

At present in the German mechanical engineering industry 1 % to

2 % of the sales volume are spent on writing offers [1]. These high
expenses are necessary, because the total production costs must be
estimated accurately at the very moment of placing the offer. Due to
an increasing number of offer demands and a decreasing transaction
rate the expenses per offer have to be reduced. So, the making of a
quotation has to be made cost-effective and thereby more economic.
Furthermore, advantages over competitors can be achieved by an
accelerated tendering.
Parallel to the planning and manufacturing of a product four phases
of calculation are carried out in an enterprise. Quotation costing (1)
helps to make an estimation of the expected manufacturing costs
and to set up the offer price. Preliminary costing (2) and & to an
increased extent ' production attending calculation (3) are used for Fig. 1: Offer Price Range for six Sets of Progressive Dies
cost control during manufacturing. Statistical cost accounting (4)
reviews the commercial success of an activity after the production. were sent to toolmaker enterprises. The analysis (fig. 1) reveals
The accuracy of the available cost information rises from quotation variation ranges in the offer prices between -29.9 %/+48.7 % for
costing to statistical cost accounting. Here, quotation costing is tool set #1 and -59.2 %/+76.2 % for tool set #3. The variations are
affected with a high risk, because the binding offer decides over the related to each average price calculated from the six offers per set.
commercial success of the future manufacturing order. Apart from the different cost structures in the individual enterprises,
The manufacturing of sheet metal parts in large scale production is the price differences were caused by the rough cost estimation. In no
mostly done with progressive dies. The design and production of case a plate sequence was developed to determine the exact tool
these tools are responsible for an essential part of the manufacturing structure and thus the costs.
costs. Because of the high outputs multiplier an accurate estimation State of the art are calculation sheets for PC-based spreadsheet
is necessary especially in large scale production. Thereby, a special programs [3]. These calculation sheets are usually specifically de-
estimation method of these tool costs must be considered. At pres- signed for the individual enterprise. The input is done manually or
ent, a plate sequence has to be developed for each exact determina- via a menu. Connections to internal production planning and control
tion of the tool costs, which requires up to four weeks time. systems and the use of existing databases are impossible. Besides
Despite the central economic importance of a binding quotation, this procedure, there are other methods of calculating sheet metal
many small and medium sized enterprises (SME) practice only a parts in which the estimation of costs is based on a partially auto-
rough estimation for the submission of a tender. This results e.g. in mated operations scheduling [4] or performed using MTM-based [5]
deviations of more than 200 % in the field of quotation costing for or consumption-based cost functions [6]. An integrated software
progressive dies, without misrepresentation of different enterprise system for offer and job order planning on the base of the hierarchi-
specific overhead rates. On the other hand only 50 % of the enter- cal similarity search was developed in the field of tool and mould
prises stated that they executed this statistical cost accounting only making [7].
occasionally or never [2]. In such cases no cost control is given.
In an inquiry, requests for quotations for six progressive die sets
Method of Feature-Based Quotation Costing
*
The authors are grateful for the financial assistance of the Industry Research
Foundation (Stiftung Industrieforschung), which supports the project work
within the framework of the research project ( Quotation Costing for the
The procedure for cost estimating introduced here is based on cost
Manufacturing of Sheet Metal Parts". features. This procedure was successfully implemented for turned


 
  
 !""#$ #%
Fig. 2: Proceeding of Feature-Oriented Quotation Costing Fig. 3: Cost Functions Based on the List of Features

parts [8]. Input data for the feature-based quotation costing (fig. 2) for the particular manufacturing method. For each feature, the most
is extracted from the customer‘s inquiry. These data include profitable manufacturing technology is selected by comparing the
workpiece material, quality of the workpiece and required batch results of the cost functions of the different manufacturing methods.
size. The result of the quotation costing is a supply price, in which A table shows the most profitable choice of the possible manufac-
the product costs are calculated and estimated with relation to the turing technologies.
selected manufacturing technique and lot size, the date of delivery
and the pricing policy.
The determination of the manufacturing costs is performed in three Costing of Progressive Dies:
steps. In the first step the CAD-model is analysed automatically Application of Artificial Neural Networks
with regard to geometric and technical features. For the automatic
recognition of the manufacturing structure of the workpiece, an An important aspect in costing is the consideration of the tool costs
existing analysis and classification system for sheet metal parts [9] for progressive dies [10]. The costs of standard tools and previously
is used. The feature recognition is carried out with analytic geomet- manufactured tools are stored in a database and are applied to the
ric analysis, semantic networks and neural networks. The output of costing. The costs of new progressive dies are estimated automati-
this system is the base for the quotation costing to be implemented. cally by the use of Artificial Neural Networks (ANN) [11]. In con-
The list of recognized features can be completed interactively. The trast to other solutions [12] the manufacturing costs of the tool set
next step is to assign the features to possible manufacturing tech- are calculated on the base of workpiece features not on base of the
niques. In this step the different features are transformed into tech- tool layout.
nique specific base units. These are machining times, profile lengths The aim of the calculation is the consumption of working hours in
or batch sizes. In the third step the base units are transformed into the toolmaker´s shop estimated by ANNs (fig. 4). The input layer is
costs. At this point the internal costing technique is applied, the formed by the information derived from the product model of the
internal key data (workcenter rates, overhead rates, etc.), cost of workpiece, not from the plate sequence. For this purpose, tools
tools and costs for refinishing, transport, and distribution flow into already subjected to statistical cost accounting were investigated in
the cost estimation. two toolmaker enterprises and analysed for use in data systems.
For the representation of sheet metal parts, a specific product model Based on the given spectrum of workpieces and the available data,
with a particular scope in the cost-orientated description of geo- the manufacturing technologies of progressive dies were limited to
metry- and manufacturing-features was defined and implemented. cutting, bending, punching and stamping. Thus, 41 tools were avail-
Various aspects are included in the product model: There are usual able.
details such as objects for components, sheet thickness, expansion For the input layer of the ANN, the following features were ex-
of the developed sheet, information on the material and the exact tracted from the CAD-Model: The number of strokes, the number of
geometrical model. On the one hand, special information on bends, tool stages, the number of border elements and stampings (which
features and joining techniques are included. In addition there are
items for information about extra tools (type and number, costs per
tool, and medium endurance) and costs for indirect functional areas,
e.g. the construction department for production facilities or the
operations scheduling.
The determination of the cost-orientated base units (fig. 3) is based
on the list of features of the piece stored in the product model. In
this list, the possible and internally existing manufacturing technolo-
gies are related to the recognized or defined features. All features
are rated as cost-oriented base units in relation to the possible manu-
facturing methods. This means that the cost drivers are derived
according to the technology. For example, in the case of profile
cutting by nibbling or laser beam cutting the processing time is
estimated depending on the profile length, numbers of holes und the
general process parameters like the velocity. Cost functions depend-
ing on the lot size are formed on the basis of these base units to-
gether with the key data of the processing equipment. The cost
function is discontinuous due to the average tool endurance given Fig. 4: Neural Networks for Cost Estimation of Progressive Dies

)* +,-./01 2-34352 366,2 357-8:97; *<!=>>)?

Fig. 5: Error Indices for the 16-15-15-6-1 Neural Network Fig. 6: Calculation Accuracy of Artificial Neural Networks

generally determine the costs). For the description of the contour specifically trained for each enterprise.
complexity the lengths of boundary and inside contours and the An approximate estimation of direct material costs for progressive
number of inside contours are used. Bends are characterized by the dies by ANNs is impractical, because the possible requirements of
number of the different bending directions, the tolerances of angles the customer for used materials and used components are too nu-
and the number of angles differentiated according to W < 90°", merous. A reliable information base is not available. For the estima-
X = 90°" and Y > 90°". The material used is described by strength and tion of direct materials, a database-based calculation sheet is used
sheet thickness. for the estimation of direct material and vendor part costs. On this
The determination of the ANN´s topology has been done in two sheet the desired components can be selected manually by the cost
steps: In the first step, the network topology was approximated by estimator.
the use of the Z FlexNet" algorithm with separated data sets for each The costing of the tools is finally achieved by the calculation of the
enterprise. The [ FlexNet" algorithm is an option of the applied ANN consumption of hours in the toolmaker´s shop, multiplied by the
toolbox \ FAST" [13] which considers the learning algorithms back- specific workcenter rate and an estimation of the material costs.
propagation, quickprop and resilient propagation. In the studies with Then, the costs per workpiece are calculated using the estimated
] FlexNet", networks with 16 input neurons and three hidden layers costs of tools and the workcenter rate of the press connected with
gave the best results under the application of the quickprop algo- the number of strokes and the costs of materials. Overhead costs are
rithm with a calculation accuracy of ca. 25% (training error index: calculated with the cost estimate sheet described below.
0.004422, validation error index: 0.40297).
In the second step, systematic variations of the number of hidden
layers and of the number of neurons per hidden layer were applied. Cost Estimate Sheet
In this investigations an ANN with three hidden layers of the topol-
ogy 16-15-15-6-1 gave the best results. A training error index of Classic job order costing as it is usually done in small and medium
0.00741855 and a validation error index of 0.21164967 were achie- sized enterprises, has the disadvantage of high overhead rates in
ved with this network (fig. 5). The validation of this network was enterprises with a high degree of automation [14]. Since the ex-
done with four workpieces. The costing results of these tools were penses spent on the introduction of a quotation costing system must
unknown to the ANN. Thereby, calculation accuracies between be kept on a low level, job order costing was retained as a basis. It
2.5 % and 10.8 % were gained (fig. 6, table 1). This accuracy was was then expanded with methods of Activity-Based Costing by
achieved on a base of 16 data sets for the training of the ANN. considering workcenter rates or resource consumption oriented cost
Finally the data sets of the two investigated enterprises were mixed functions to manufacturing steps and to the manufacturing of fea-
together. The same two steps for the determination of an optimized tures. Workcenter rates and cost functions are taken into account as
network structure were executed. The training results with valida- full costing rates and variable costing rates. The different rates are
tion errors of more then 1.0 (fig. 7) show that the networks cannot used for a comparison between fixed and variable costs for a sheet
be trained on the mixed data sets. A comparison between the inves- metal part. This dual calculation strategy allows a simple determina
tigated toolmaker‘s shops showed a lot of differences in methods
and organization structures. The networks are not able to generalize
between the differences of the two enterprises. They have to be

Network Layer Topology Minimum Maximum Average

No. of the ANN Deviation Deviation Deviation

1 16-15-15-6-1 2.5 % 10.8 % 6.7 %

2 16-15-15-9-1 0.5 % 21.3 % 7.4 %

3 16-3-1 2.5 % -30.9 % 10.5 %

4 16-18-18-9-1 1.5 % 37.5 % 10.5 %

5 16-18-18-9-1 0.2 % -34.9 % 15.4 %

Fig. 7: Training Results of Artificial Neural Networks with Data Sets from
Table 1: Network Verification with four Unknown Tool Sets two Enterprises (FlexNet-Algorithm)

@ABCDEF G BHIHJG HKKAGHJLBMNLO PQ!RSSTU TV

 Conclusions

A feature oriented method for the quotation costing of sheet metal

parts is described and implemented in a software system. By means
of this system, it will become possible to calculate offers quickly, to
make costing transparent and to consider enterprise-specific key
data. Therefore, a cost orientated product model was developed and
implemented in which geometric features are transformed into
costing features. With the aid of these costing features, the direct
costs can be deducted.
Cost calculation is based on job order costing, expanded with the
workcenter costing. Workcenter rates are taken into account with
full costing rates and variable costing rates. Based upon this meth-
ods of contribution costing are integrated in the cost estimate sheet
Fig. 8: Database Structure for Offer Management to support the cost estimator calculating the profit contribution of
an order.
tion of the profit aimed at as well as the contribution margin for a The estimation of tool costs for progressive dies is based on the use
manufacturing task using methods of contribution costing. of enterprise-specific trained Artificial Neural Networks. The neces-
The starting points for the costing are the order in combination with sary data are extracted from the workpiece not from the tool layout.
the product model including the lot size and the number of subcon- This cost determination is done without the generation of a plate
tracts. In the cost estimate sheet the costs for each order (e.g. costs sequence. With this method the workcenter costs in the toolmaker‘s
for the construction of the production facilities, operations schedul- shop can be calculated. Direct material costs and costs for supplied
ing or extra tools) are considered once in the costing. The number of parts are considered by the use of a database. Within the investi-
subcontracts forms the factor for the consideration of the costs per gated range of workpieces calculation accuracies between 5 % and
subcontract (e.g. for change over of tools) and the distribution ex- 15 % can be achieved.
penses. Finally for each workpiece, the material costs including the These introduced methods are implemented in a quotation costing
waste, the manufacturing costs across all manufacturing steps, the system for personal computers with the operating system MS-Win-
costs for finishing and transport are considered. dows. The standards SQL and ODBC are used as database inter-
To increase the accuracy of costing, the system also includes the faces, so that different database systems may be applied in the sys-
costs for the internal transport and the transport between different tem.
branches. The costs for heat treatments and surface treatments of the
workpiece are considered in special cost functions. These are di-
vided into weight-depending costs (e.g. for heat treatment) and area- References
depending costs (e.g. for purifying and coating). 1 Bronner, A.: Angebots- und Projektkalkulation - Leitfaden für tech-
nische Betriebe. Springer, Berlin Heidelberg 1996
2 Weber, J.: Kostenrechnung im Mittelstand. Schäffer Poeschel, Stuttgart
Database Structure for Offer Management 1992
3 Bopp, R.: Aufwand verringern. Industrie-Anzeiger 117 (1995) 17,
For the management of the data needed for quotation costing a S. 60-61
special database structure (fig. 8) was developed and implemented. 4 Wierda, L. S.: Cost Information Tool for Designers. Dissertation TU
The structure is divided into two parts: an enterprise specific and an Delft (NL) 1990
offer/order specific section. 5 Boothroyd, G., Dewhurst, P.: Sheet Metalworking: Material Selection
and Process Costing. Wakefield 1990
The enterprise specific part primarily contains the key data of the
enterprise, e.g. overhead rates for different departments or default 6 Kümper, R.: Ein Kostenmodell zur verursachungsgerechten Vorkalkula-
tion. Dissertation RWTH Aachen 1996
values for the pricing policy. Additionally in the workcenter data-
7 Brunkhorst, U.: Integrierte Angebots- und Auftragsplanung im
base the workcenter rate (variable costing/full costing), information Werkzeug- und Formenbau. Dissertation University of Hannover 1995
about technical restrictions (volume of working area, maximum
8 Wolfram, M.: Feature-basiertes Konstruieren und Kalkulieren, Disserta-
workpiece weight) and machining times (cycle, change-over and tion TU München 1994
handling time) are stored. 9 Geiger, M., Greska, W.: Analysis and Classification of Sheet Metal
The offer specific data are distributed over several small databases Components. Production Engineering 1 (1993) 1, S. 191 - 196
to avoid redundant information in the storage. The customer data- 10 Malstrom, E. M.: Manufacturing Cost Engineering Handbook. Bd. 5
base contains usual administration information as well as a proposal Marcel Dekker, New York 1984
for a discount and information about solved manufacturing tasks. 11 Haykin, S.: Neural Networks. Macmillan College Publishing Company,
Within the workpiece database geometric information and cost New York 1994
information of statistical cost accounted products is managed. The 12 Krannich, M.: Tool Cost Calculation using Neural Networks. In: Kals,
handling of cost driving features of workpieces (e.g. recognized H. J. J., Geiger, M., Shirvani, B., Singh, U. P.: Sheet Metal 1998 -
Proceedings of the 6th International Conference. Twente 1998, S. 221-
form features, contour length) is done in a special feature database. 230
Arras, M. K., Mohraz, D.: FAST v2.2 u FORWISS Artificial Neural
Additionally there are special databases for given tenders (e.g. date 13
of delivery, batch size or price), materials (e.g. costs, availability, Network Simulation Toolbox. Erlangen 1996
strength) and special tools (costs or average wear). All databases are 14 Cooper, R..: Activity-Based Costing. Kostenrechnungspraxis (1990) 4,
linked together in a relational database model for an effective use. S. 210-220

^_ `abcdef gbhihjg hkkag hjlbm:nlo pq!rss^t