Available online at www.sciencedirect.

com

ScienceDirect
Available online at www.sciencedirect.com

ScienceDirect
Procedia Manufacturing 00 (2019) 000–000
www.elsevier.com/locate/procedia

Procedia Manufacturing 42 (2020) 351–355

International Conference on Industry 4.0 and Smart Manufacturing (ISM 2019)

An integrated approach for power transformer modeling and manufacturing

Christian Lettnera, Michael Mosera, Josef Pichlerb*
a
Software Competence Center Hagenberg GmbH, Softwarepark 21, 4232 Hagenberg, Austria
b
University of Applied Sciences Upper Austria, Softwarepark 11, 4232 Hagenberg, Austria
* Corresponding author. Tel.: +43-50804-22039 ; fax: +43-50804-21599. E-mail address: [email protected]

Abstract

Essential characteristics of smart factories, such as flexibility and resource efficiency, can be leveraged and improved by the power of machine
learning and optimization techniques. For instance, the manufacturing process of a power transformer core constitutes a highly complex
optimization problem. It involves creating a cost optimal slitting plan that meets all customer requirements and at the same time takes into account
flexible and short-term constraints from production (e.g. current available metal bands in stock). As many of these constraints rely on forecasts,
a learning system may provide the necessary predictions for these constraints. In addition, companies apply and maintain engineering software
for a variety of tasks in construction, simulation, and interpretation of data. For instance, electrical engineers use a variety of tools to design an
initial model of a power transformer according to customer requirements and constraints. Such tools often incorporate knowledge that serves as
input for optimization and forecast models as described before. If these models are improved over time using external machine learning libraries,
the newly developed models must find their way back into the implementation of engineering tools. Knowledge scattered across multiple software
systems bears risk of being inconsistent. Furthermore, keeping knowledge consistent without a systematic approach is time-consuming and error-
prone. In this paper, we describe an approach that leverages software engineering methods and tools and that supports knowledge transfer between
software systems for optimization and modelling tasks. The approach follows the idea of a single source of knowledge together with
transformation into different representations, as required by different engineering tasks. The proposed approach was introduced at an industrial
site to improve the manufacturing process of power transformer cores.

© 2020
© 2020The TheAuthors.
Authors. Published
Published by Elsevier
by Elsevier B.V. B.V.
This is an open access article under the CC
This is an open access article under the CC BY-NC-NDBY-NC-ND
licenselicense (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-reviewunder
Peer-review underresponsibility
responsibilityofof
thethe scientific
scientific committee
committee of the
of the International
International Conference
Conference on Industry
on Industry 4.0Smart
4.0 and and Smart Manufacturing.
Manufacturing.

Keywords: cutting stock problem, manufacturing, optimization, machine learning, software engineering.

1. Introduction aim of previous research [1]. To optimize the production of

such a core, which consists of hundreds of individual metal
Power transformers are strategic nodes in every electrical sheet layers, a system was developed that combines machine
power supply network and there is a competitive international learning based predictions for critical quality attributes and a
market for producing these critical infrastructures. Decades of fast and efficient optimization of material usage. Based on the
optimization have resulted in transformers being extremely quality of the raw materials (transformer metal sheets) and
energy efficient. However, eco-design measures have forced an geometric features, the machine learning model is able to
even closer look at the losses produced. In recent years, strict predict the resulting product quality in terms of dissipation
noise control regulations and client specifications have made losses and noise generation.
the acoustic performance a major topic for transformer The aim of machine learning based prediction is to improve
manufacturers. existing prediction formulas (which incorporate both expert
Optimizing the losses within the core and the resulting sound knowledge as well as prior trainings) with adjustment factors
levels under various guaranteed measuring conditions while that will reduce the prediction error. Besides expert knowledge
still producing the transformer at competitive prices was the and prior trainings, the main source of prediction formulas is
2351-9789 © 2020 The Authors. Published by Elsevier B.V.
2351-9789
This is an©open
2020 The Authors.
access Published
article under bytheElsevier B.V.
CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review
This is an openunder
access responsibility
article under the of
CCthe scientificlicense
BY-NC-ND committee of the International Conference on Industry 4.0 and Smart Manufacturing.
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the International Conference on Industry 4.0 and Smart Manufacturing.
10.1016/j.promfg.2020.02.076
352 Christian Lettner et al. / Procedia Manufacturing 42 (2020) 351–355
2 Author name / Procedia Manufacturing 00 (2019) 000–000

the source code (e.g. Fortran) of design software that engineers model. The resulting model is also used in simulation for further
use to construct an initial model of a power transformer (iterative) optimization.
according to customer requirements and constraints.
Unfortunately, the computation of these formulas is often
highly optimized in terms of running time or memory usage,
which makes it difficult to comprehend and extract the required
knowledge directly from the program source code [2, 3]. For
the same reasons, the source code often is not structured
according to the different types of formulas, which further
complicates the understanding of formulas required for the
machine learning based prediction. Likewise, the reworking of
the improved formulas in the source code is laborious and error-
prone.
The formulas actually implemented in the design software
and improved by machine learning prediction have a significant
impact on quality characteristics of the final product (which
cannot be changed after manufacturing). Therefore, a principal
engineer (often without programming expertise) must approve Fig. 1. Power transformer modeling and manufacturing.
this type of knowledge. For the same reasons given above, the
formula must be made available to the engineers in a readable Produce. After the results of the simulation are satisfactory,
(mathematical) notation. the manufacturing of the power transformer begins with the
Keeping the knowledge (formula) consistent between the selection of coils (according to core sheet design) and the
design software, the machine learning framework, and the cutting of coils into core sheets. These two steps are highly
principal engineers while supporting different representations optimized by the machine learning approach presented in this
is the aim of the approach described in this paper. In Section 2 paper.
we describe the manufacturing of power transformers and the Measure. After production, the final power transformer is
challenges involved together with an overview of the approach measured under various operating conditions (e.g. full load
exemplified at an industrial site. Section 3 briefly describes the operation, short-circuit operation) for mechanical and electrical
machine learning based prediction for optimization of material properties including noise etc. The product is finished with this
usage in power transformer manufacturing. In Section 4 we step and can be delivered. From the producer's point of view,
describe the software engineering methods and tools developed however, another important step will follow. The measured
to keep knowledge consistent across different representations values are compared with the original requirements and the
and tasks. In Section 5 we conclude with the results and comparison serves to optimize the modelling.
experiences from the application of our approach. Optimize. Data from measurement serve as input for the
subsequent optimization step. For instance, the material usage
2. Background and Overview optimization needs all related data, e.g. raw material attributes
(measured by a dedicated measurement system), purchase data,
2.1. Power Transformer Design and Manufacturing and quality assurance data together with an already existing
prediction formula. This formula is considered for generating a
Fig. 1 shows a typical sequence of activities involved in the base prediction given the specifics of the order data of a power
modeling and production of power transformers. The activities transformer. This formula is then adjusted by the learning
model, produce, measure, and optimize constitute a cycle of schema with a prediction error by means of adaptive filter that
continuous quality improvement on the manufactured product. optimally adjusts the unknown parameters of the function.
In the following, we describe the individual activities as well as
the overall effect together with challenges that results from the
interaction of these activities. 2.2. Challenges
Modeling. First, electrical engineers calculate an optimal
design of an electrical machine that meets the customer The above-mentioned existing prediction formulas are
requirements and minimizes the production costs as far as typically implemented in a design software and therefore
possible without compromising the quality of the power intensively used in the design of power transformers. Because
transformer. As power transformers are individual products that of this relationship, it is very important that the current version
have to be designed from scratch each time, the complex design of the formulas (used in the design software) are used as the
and modeling can no longer be done by hand. Today, engineers starting point for optimization, and above all that the formulas
use a huge set of software tools, where every software tool resulting from optimization find their way back into the design
covers a specific aspect in the design (e.g. temperature losses in software. The electrical engineer uses the design software
a winding or core sheet losses) and either automatically creates improved from the current “lifecycle” to design the next power
an initial model based on customer requirements or verifies transformer.
geometrical, mechanical, or electrical properties of the resulting The adjustment of formulas in the design software after
optimization means changing the source code, testing and
deploying the changed software so that engineers can use it.
 Christian Lettner et al. / Procedia Manufacturing 42 (2020) 351–355 353
Author name / Procedia Manufacturing 00 (2019) 000–000 3

software. The data-driven approach chosen for this task ensures

that for example changes in raw material properties are detected
and will be represented by an improved formula. All
improvements on the formula must be carried out in such a way
that the overall formula remains interpretable and
comprehensible, so that it can be implemented in the design
software and verified by technical experts. The formulas are
trained using order data and measurement data.
To preserve the interpretable and comprehensible formulas,
existing factors in the formula can be optimized or new factors
Fig. 2. Integration of optimization/learning with design/modeling tasks.
may be added to the initial formula based on training data
generated from measurements performed on recently produced
Besides the source code, the software developer must update transformers.
the documentation, because principal engineers must sign the Fig. 3 shows the learning architecture used in the proposed
formula used in the design software. Keeping these documents approach. The learning scheme accepts as inputs order data
up-to-date manually is not only time-consuming, but also error- including the properties of the material sheets, and
prone in terms of completeness and details. measurement data, which includes measurements of the
property under investigation. The goal is to be able to improve
2.3. Approach Overview already existing prediction formulas (which incorporate expert
knowledge as well as prior trainings) with adjustment factors f
Fig. 2 shows the overall approach that integrates the that will reduce the predication error. In particular, the sequence
optimization and learning tasks with the design and modeling of computations is as follows [1]:
tasks. A principal design decision of our approach is that the 1. An already existing prediction formula is considered for
design software (used by engineers) is the single source of generating our base predictions ψ_k (given the specifics
knowledge and other representations will be derived from the of the order data of a transformer). Such predictions
source code. We developed program analysis methods with tool may incorporate prior expert knowledge as well as prior
support to analyze the source code of the design software and trainings.
to transform extracted knowledge (e.g. formulas) into 2. We introduce an adjustment factor to the current
mathematical / engineering style of documentation. For the predictions (summarized within function 𝑓𝑓). The
principal engineer, our tool generates ready-to-use function 𝑓𝑓 is usually defined as a polynomial of
documentation in the MS Word format. The extracted set of continuous variables of a transformer (e.g., its mass). It
formulae are not only suitable for principle engineers but also may also accommodate dependence on discrete
serve as initial formulae for the machine- learning framework. variables of a transformer by simply considering
different polynomials for different values of the discrete
Based on the initial formula and other input parameters the variable. The parameters of such polynomials need to
machine learning based prediction delivers optimized formulas, be estimated.
which serves two purposes. First, the optimization of the 3. Given the prediction error ek of the adjusted formula, we
material usage takes this output to predict product quality in introduce an adaptive filter, which optimally adjusts (in
terms of dissipation losses and noise generation. Second, the a way to be determined) the unknown parameters of the
software engineers update the source code of the design function f. A schematic of the adaptive adjustment of the
software. As shown in Fig. 2 the knowledge flows from the parameters is shown in Fig. 3.
design software to documentation and ML-based prediction The adaptive filter chosen for this purpose is exponentially-
and, in updated form, back into the design software. After one weighted regularized least squares, implemented in a recursive
such cycle, the design software is again the single reliable manner. Note that in the absence of any existing prior prediction
source of knowledge. Our code analysis tool, afterwards, formulas, the above setup can be used to generate prediction
automatically extracts this knowledge and provides different models from scratch.
representations.

3. Optimization and Learning

3.1. Machine Learning based Prediction

Prediction formulas relate desired transformer properties to

the properties of the metal sheets selected to build it. The
objective of the learning framework [1] is to continuously
improve these prediction formulas extracted from the design Fig. 3. Learning architecture (see [1]).
354 Christian Lettner et al. / Procedia Manufacturing 42 (2020) 351–355
4 Author name / Procedia Manufacturing 00 (2019) 000–000

3.2. Optimization of Material Usage optimization tasks and must also be updated based on
the optimization results.
The optimization of material usage [1] uses prediction For all scenarios mentioned, it is important that the
formulas to perform an overall optimization to get a cost documentation actually describes the knowledge (e.g. formula,
optimal slitting for the metal sheets. It is able to create optimal decision tables, and parameter) that is implemented in the
slitting plans of a set of available metal coils into bands of design software and therefore actually used during the design
desirable width, and add customer constraints of properties of process. The co-existence of source code and documentation
the final product. It provides flexibility with respect to the implicates several challenges. Maintaining both documentation
constraints and objectives. and source code independently results in high writing effort and
The problem can be mathematically formulated as a multi- in a documentation that is likely to be inconsistent by itself and
objective optimization problem and can be thought as a inconsistent with respect to the documented source code.
generalization of the well-known cutting-stock problem. The Documentation may be out-of-date or non-existent at all for
main objectives are: reducing production time, reducing legacy code maintained and evolved over several years.
machine adjustments, and optimizing physical transformer To tackle these problems, our approach leverages methods
properties. The constraints can be grouped into the following from static code analysis to generate documentation from
categories: job constraints, machine constraints, and physical source code of the design software. Our approach (named
constraints. Rulebook Generator, or RbG for short) accepts the source code
Physical properties must be either fulfilled in any case or are of the design software, applies static program analysis to
subject to optimization. Accordingly, they are modelled as a transform program statements into mathematical formula and
constraint or in the objective function. Physical transformer decision tables, and finally applies transformation rules to
properties directly depend on the quality of the used metal coils generate ready-to-use documents. To accomplish this process,
for production. This dependency results in a non-linear RbG integrates following mechanisms.
objective function and non-linear constraints. The optimization 1. RbG processes source code annotations to identify
of material usage is implemented as a stochastic-local search software components that actually contain relevant
algorithm, namely a genetic algorithm, with a diversification knowledge. Software developers provide these
strategy to avoid local minima. annotations for “entry points” of computations and on
this basis, RbG uses static program analysis [5] to
3.3. Model Management determine all the components involved. Even
knowledge is scattered across different software
The continuous and incremental improvement of the models components only 5-10% thereof are of actual interest
(formulas) results in a large number of hierarchically [3].
dependent models. To manage these models efficiently, a 2. RbG processes source code statements and resolves
model management system [4] has been implemented, that logs unstructured control-flow (e.g. break/continue
and manages important configuration parameters and results of statements within loop statements, or goto statement) by
the learning framework in a traceable manner. means of refactoring resulting in well-structured
control-flow, i.e. single entry/exit points of statements
4. Maintenance and Documentation of Design Software [6]. Next, RbG leverages symbolic execution [6] to
eliminate intermediate variables and constructs
The design software used in the modeling step is an example assignment statements that only contain input/output
of software systems applied for a variety of tasks in parameters of the analyzed algorithms. RbG
construction, simulation, and interpretation of data. subsequently transforms such constructs into
From a software engineering perspective, maintenance and mathematical formulae together with the condition
evolution of these software systems requires extensive (extracted from the path condition) under which the
documentation of domain knowledge, including physical and formula is applied. As different execution paths (with
mathematical description of the problem together with selected distinct path conditions) may result in multiple
numeric algorithms to solve the problems. This kind of formulas, the general output is a decision table with
documentation is created and updated over time for several formulas and their conditions.
reasons. 3. Further, RbG provides an interpreter to visualize
1. Software developers and domain experts (e.g. electrical extracted knowledge exemplified by input data in form
engineers) use this documentation, in particular for of function plots.
infrequent changes of the software as well as to 4. Users can control the extraction / generation process
introduce novices. with a set of further annotations as well as enrich the
2. A principal engineer relies on the same documentation extracted content by text and images already during
for the approval of the design software. This step is part generation. In this way, the generated documents are
of an overall quality process to ensure that engineers use ready-to-use and need not be changed further.
latest findings from research and optimization as
described in Section 3. Fig. 4 shows the overall software architecture of the RbG
3. Last and in the focus of this paper, certain formulas approach with information flow from left to right.
contained in the design software serve as input for
 Christian Lettner et al. / Procedia Manufacturing 42 (2020) 351–355 355
Author name / Procedia Manufacturing 00 (2019) 000–000 5

components. Even if developers are familiar with source code,

they benefit from the immediate alternative representation. The
mathematical representation and especially the function plot
diagrams make it easier for developers to control the
implemented formulas. Second, RbG generates documentation
whenever the build server checks out changed source code. In
this way, the build server produces new version of a software
together with documentation. Third, data scientist use
generated documentation as initial input for machine learning
based prediction. For optimization of the material usage at an
industrial site, the initial formulas were extracted and
documented by the RbG tool and, after optimization, the
software developers updated the source code accordingly and
Fig. 4. Documentation generation architecture. verified the changes made again using the generated
documentation.
In the first step, RbG parses source code of various
programming languages (e.g. Fortran, C, C++ which are still Acknowledgements
predominant in engineering domains [2, 3]) and constructs an
intermediate representation, the so called Abstract Syntax Tree The research reported in this paper has been supported by
(AST) [7]. A key design decision here is that the AST is the Austrian Ministry for Transport, Innovation and
independent of the programming language parsed in the Technology, the Federal Ministry of Science, Research and
previous step. In this way, all subsequent analysis and Economy, and the Province of Upper Austria in the frame of
transformation steps can be applied in a uniform way. In other the COMET center SCCH.
words, RbG can easily be enhanced by a further programming
language. References

5. Results and Discussion [1] Chasparis C, Zellinger W, Haunschmid V, Riedenbauer M, Stumptner R.

On the optimization of material usage in power transformer manufacturing.
Data analytics, machine learning and artificial intelligence In the Proceedings of the 8th International Conference on Intelligent
are at the core of Industry 4.0 (the smart factory). Significant Systems, 2016. p. 680-685.
[2] Moser M, Pichler J, Fleck G, Witlatschil M. RbG: A documentation
aspects of a smart factory, e.g. flexibility and resource generator for scientific and engineering software. In Y. Guéhéneuc, B.
efficiency, can be improved by leveraging the power of Adams, A. Serebrenik (editors), Proceedings of the 22nd IEEE
machine learning and optimization techniques. As a result, International Conference on Software Analysis, Evolution (SANER'15),
more and more knowledge is created, implemented in different 2015, p. 464-468.
software systems and has to be changed over time. The [3] Pichler J. Extraction of documentation from fortran 90 source code: an
industrial experience. In A. Cleve, F. Ricca, M. Cerioli (editors),
approach presented in this paper was developed to keep the Processdings to the 17th European Conference on Software Maintenance
distributed knowledge about different software systems and and Reengineering (CSMR 2013), 2013, p. 99-104.
about different phases of modeling and optimization [4] Polyzotis N, Roy S, Whang SE, Zinkevich M. Data lifecycle challenges in
synchronized. The approach was introduced at an industrial site production machine learning: a survey. SIGMOD Rec. 47, 2 (December
and applied in a project for optimization of material usage. The 2018), 2018. p. 17-28.
[5] Nielson F, Nielson HR, Hankin C. Principles of program analysis. Springer,
RbG tool was applied to a dozen of design software systems to Dec. 2010.
generate up-to-date documentation. The tool is seamlessly [6] King JC. Symbolic execution and program testing. Commun. ACM, vol.
integrated with the software development toolchain for 19, no. 7, Jul. 1976, pp. 385–394.
developers (MS Visual Studio) as well as on the build server [7] OMG. Architecture-driven modernization: abstract syntax tree metamodel
and is used in several scenarios. First, software developers use (ASTM), version 1.0. OMG. 2011.
RbG-based generation of documents and in particular diagrams
during development / maintenance of certain software