Application of Artificial Intelligence

Technology in Product Design

 Naoyuki Nozaki  Eiichi Konno  Mitsuru Sato  Makoto Sakairi

 Toshiyuki Shibuya  Yuuji Kanazawa  Serban Georgescu

Artificial intelligence (AI) technology applied to product design in Monozukuri (Japanese

way of manufacturing) aims to provide computerized support to various tasks in developing
products that currently rely on human experience. As the conventional approach, in which
knowledge and rules are explicitly given, has its limit, new technologies based on machine
learning have been recognized as important in research and development of AI. Applying
machine learning that predicts data to be acquired in the future with a certain accuracy, we
can obtain efficient and less-variable judgment and eliminate conventional work depended
on personal knowledge or experience. In order to apply AI technology to a product develop-
ment environment, MONOZUKURI AI framework was developed on the cloud as a system to
facilitate efficient collection of product development data, and, at the same time, for manag-
ing and leveraging learning models extracted from such data. By connecting this framework
with Flexible Technical Computing Platform (FTCP), Fujitsu’s integrated development platform,
this new design development environment will provide various design tools on a platform with
new, enhanced design-assisting features. This paper describes various cases in which machine
learning is applied to designing, and presents our plan to introduce it into an integrated devel-
opment platform.

1. Introduction has been promoting AI systematization through its

The traditional approach to solving a given prob- concept, Human Centric AI Zinrai.1) Through the auto-
lem with software has been to clearly formulate a matic acquisition of knowledge and rules from data, AI
procedure, i.e., an algorithm. However, there are many technology has the potential to overcome the limits of
problems that do not lend themselves to the clear for- the conventional approach based on the presentation
mulation of an algorithm. In the case of estimating of explicit knowledge and rules. Fujitsu has accumu-
the number of layers in a printed circuit board (PCB), lated an enormous body of product design data from
described later on in this paper, a skilled designer is its Monozukuri (Japanese way of manufacturing) activi-
able to judge the number of layers based on past ex- ties to date, and the successful modeling for learning of
perience rather than consciousness of a set of clearly the knowledge data included in all this product design
defined rules. Product development involves many data would allow the application of this knowledge and
such tasks that require designers to make judgments these rules to new designs. To this end, the authors
based on experience. There are many such problems are evaluating the effectiveness of machine learning in
where the clear formulation of an algorithm is difficult product design, and are working on applying machine
and where the incorporation of all the factors that need learning to the product design environment.
to be considered requires complex calculations, which This paper presents cases of the application of
is also difficult. machine learning to product development. Next, it
Meanwhile, the importance of machine learn- introduces Fujitsu’s MONOZUKURI AI framework as an
ing, an artificial intelligence (AI) technology for the example of the incorporation of machine learning in
utilization of the big data generated by the Internet the Flexible Technical Computing Platform (FTCP), an
of Things (IoT), has been drawing attention. Fujitsu integrated development platform.2) Finally, it describes

FUJITSU Sci. Tech. J., Vol. 53, No. 4, pp. 43–51 (July 2017) 43
N. Nozaki et al.: Application of Artificial Intelligence Technology in Product Design

Fujitsu’s aims for its Human Centric AI Zinrai system for to optimize the number of layers, it is common to route
the systematization of AI in the design and develop- key signals (such as high-speed signals) temporarily
ment environment. and estimate the number of layers in a process that can
take several hundred hours.
2. Examples of application to electrical/ Thus, the task of estimating the number of layers
structural design in PCBs is an important part of the PCB design process,
As examples of the application of machine learn- and it is a difficult and time-consuming task even for
ing technology in the design field, this section describes skilled designers.
the development of a function to estimate the number 2) Utilization of machine learning technology for the
of layers in PCBs in the electrical design environment, estimation of the number of layers
and the development of an automatic detection func- As described above, one method for estimating
tion of 3D model components using shape recognition the number of layers in a PCB is to provisionally arrange
in the structural design environment. the main components and perform provisional routing
of the signals and the like to which particularly severe
2.1 Application example in the electrical mounting conditions apply. However, this requires
design environment manual work and sophisticated judgment, making the
This section introduces the case of the application automation of such work an inadvisable proposition.
of machine learning to the development of a function Therefore, the authors hypothesized that it might be
to estimate the number of layers in PCBs. possible to create a model to predict the number of
1) Importance of estimating the number of PCB layers in a PCB through machine learning from past
layers product design data.
The design process of PCBs consists mainly of the Generally, the PCB of a server or a network device
circuit design stage and the board design stage. In consists of signal layers, a power supply layer, and
the circuit design stage, appropriate components are a ground (GND) layer. In the case of manual design,
selected and logical connections between the various often the number of signal layers is first estimated
components and their mounting conditions (signal empirically by the designer. As in the case of manual
line lengths, etc.) are specified. In the board design design, the signal layers were first targeted for the
stage, components and logical connections are physi- estimation of the number of layers through machine
cally arranged and wired on the board according to the learning.
mounting conditions specified in the circuit design. In the machine learning method, support vector
An important design decision to be made when regression,3) which is a regression analysis method
transitioning from the circuit design stage to the board used in supervised learning, was employed. Normally,
design stage is the number of layers in the PCB. The because the number of layers has to be estimated
task of estimating this number requires taking into con- before starting board design, the feature vectors (nu-
sideration a number of factors such as the placement merical data) that are input as training data for the
location of the various components, the wiring route, creation of the learning model are those obtained at
and the allowable size of the PCB, which is restricted by the circuit design stage only.
the target apparatus size. This is because the number Next, we introduce the initiatives that led to
of layers in the PCB affects the number of components improvement of the accuracy of the estimation of the
and the number of wiring paths, which directly affects number of layers during the learning model creation
the manufacturing cost, making this an important deci- process.
sion for eliminating design rework. The first one was to select only items close to
For this reason, highly skilled designers often pro- the estimation target product from the design data as
ceed while considering the requirements of products, training data. The target learning data was narrowed
the circuit configuration, and so on, and trade off the down by sorting past design data by product type and
various elements involved, comprehensively judging PCB type. This sorting made it possible to improve
their relative priority levels. Especially when one wants the accuracy of estimation of the number of layers for

44 FUJITSU Sci. Tech. J., Vol. 53, No. 4 (July 2017)

 N. Nozaki et al.: Application of Artificial Intelligence Technology in Product Design

specific product categories. for the automatic detection of 3D model components

The second initiative was to review the feature using shape recognition.
vectors to be input as training data. During the learn- 1) Design verification during structural design
ing model creation process, we surveyed PCB designers Structural design work and design reviews are
about the items that they emphasize for layer number usually carried out by referring to check lists that have
estimation, and we conducted a review of the feature been compiled from the accumulated experience and
vectors accordingly. know-how of highly experienced designers designing
The third initiative was algorithm optimization similar products. However, as the accumulated experi-
during learning model creation. For example, we were ence and know-how increase, so does the verification
able to improve estimation accuracy by changing the work time and the load on the designers. Moreover,
kernel function for support vector regression from linear in structural design, shapes are modeled to reflect
to nonlinear, and by changing the parameters supplied the intentions of the designer (e.g.: adding a rib for
to the kernel function. greater strength). However, the generated 3D models
Figure 1 shows the results of estimation of the are merely shape information, and computers cannot
number of signal layers based on cross validation for a judge the intentions of designers. Therefore, the only
given product type. The x axis shows the actual num- way to verify whether a design intent works correctly
ber of signal layers and the y axis shows the estimated was for a human to visually judge the location to be
number of signal layers obtained from the developed verified. As a result, as well as long verification times,
learning model. Each plot corresponds to one design there is also the issue of verification omissions due to
data. All data are plotted along the y = x line, indicat- human factors.
ing that the number of signal layers can be estimated 2) Utilization of machine learning for automatic de-
with high accuracy. tection of 3D model components
Based on these results, we plan to also estimate We investigated whether it is possible to achieve
the number of layers other than signal layers, and de- machine learning based on past design data for the
termine the layer configuration. automatic detection of similarly shaped components
(such as screws). In doing so, based on the hypothesis
2.2 Application example in the structural that similarity of 3D shapes means also similarity of 2D
design environment shapes, we tested whether image recognition using 2D
This section introduces the case of the application images can be applied to the detection of 3D shapes.
of machine learning to the development of a function The automatic detection procedure is outlined below.
(i) Twenty 2D images are generated for each com-
y
ponent from past design data (as shown in
Figure 2, the vertex of a 20-sided polygon relative
to the center of the component is taken as the
Estimated number of signal layers

camera viewpoint).
(ii) The feature vectors of each component are ex-
tracted from the 2D images.
(iii) Twenty 2D images are generated in the same way
for components for which similar shaped compo-
nents are to be detected.
(iv) The feature vectors of (iii) are extracted, their
similarity with the feature vectors stored in the
created database is calculated, and similarly
x
Actual number of signal layers shaped components are detected in the database.
This time, we created a learning model based on
Figure 1 approximately 8,000 components for which we had
Evaluation of estimation accuracy of the
number of signal layers. past design data, and we verified whether automatic

FUJITSU Sci. Tech. J., Vol. 53, No. 4 (July 2017) 45

N. Nozaki et al.: Application of Artifi cial Intelligence Technology in Product Design

Camera

Figure 2
2D image representation.

detection of screws was possible. MONOZUKURI AI framework introduced in this section.

Figure 3 shows the results of initial verification
in descending order of similarity. Approximately half 3.1 Learning model development
of the 60 components recognized as most similar con- framework
sisted of components other than screws. This was due The utilization of machine learning technology al-
to the fact that, in the case of symmetrical components lows us to predict what will occur next from new data
such as screws, a number of the 2D images generated and classify such new data based on its characteristics.
for the extraction of feature vectors in Step (i) above However, building a highly accurate learning model ap-
are actually identical, which seems in turn to reduce plicable to design requires optimization of the learning
shape recognition accuracy. We solved this problem, model by various means. To this end, as regards the
generating 2D images by changing the angle of view, work required to develop practical learning models,
so as to prevent the duplication of 2D images. In addi- we are currently developing guidelines and tools and
tion, we incorporated also component size and surface building a learning model development framework
area information. As a result, as shown in Figure 4, that will support service development using machine
the detection accuracy of similar shapes was greatly learning technology. As an example, Figure 5 shows
improved compared with the initial verification results. the development flow of a learning model based on
Thus, through the above-described initiatives, we the supervised learning method for machine learning.
were able to verify the feasibility of automatic detection In order to use the supervised learning method,
of components based on a 3D model. In the future, it is necessary to first collect training data. To improve
we plan to further improve the detection accuracy and the accuracy of the learning model, it is effective to re-
achieve automation of the verification of components. move training data that will just be noise (for example,
Further, since there are many other items to be verified incomplete design data) and classify the retained train-
besides components, including the specific shapes of ing data by applicable theme. To make the selection
parts of components such as ribs and bosses, we would of training data more efficient, we have developed and
like to also apply the obtained findings to the detection utilized tools to visualize the characteristics of design
of the shapes of parts of components. data in the form of graphs. After the training data has
been collected, it is necessary to extract feature vectors
3. MONOZUKURI AI framework (numerical data) from the training data for use as input
In applying machine learning technology to the data for the learning model.
field of product design, there are processes common An effective way to make feature-vector extraction
to various design fields and application themes. As more efficient is to expand the types of numerical data
the basis for standardizing common processes related that can be extracted from design data. The kinds of
to machine learning, we are working on a framework numerical data that are appropriate to use as feature
for utilization and development in the form of the vectors differ according to the theme to which machine

46 FUJITSU Sci. Tech. J., Vol. 53, No. 4 (July 2017)

 N. Nozaki et al.: Application of Artifi cial Intelligence Technology in Product Design

Figure 3
Initial verification results of screw component detection.

Figure 4
Improvement results of screw component detection.

Data collection and review

learning technology is applied. However, there are also
commonly used feature vectors, so increasing the types
Rework No. 1
of numerical data that can be extracted from the de-
Creation of training data
sign data is effective for developing a large number of
Rework No. 2 learning models.
Creation of learning model The final step of the supervised learning method
Rework No. 3 consists in analyzing the learning model created from
Analysis of learning model the training data and improving its accuracy. To this
end, we apply methods for improving the accuracy of
general learning models, such as comparison of predic-
Completion of highly accurate
machine learning model tion results by multiple algorithms and cross validation,
as tools for supporting work.
Figure 5 In this way, by preparing a series of environments
Machine learning flow. necessary for learning model development, it is pos-
sible to minimize rework shown in Figure 5 (rework
No. 1, 2, 3) to make learning model development more
efficient. Going forward, we plan to deal not only with

FUJITSU Sci. Tech. J., Vol. 53, No. 4 (July 2017) 47

N. Nozaki et al.: Application of Artifi cial Intelligence Technology in Product Design

supervised learning but also methods such as unsuper- functions provided by the utilization framework through
vised learning and deep learning. the web API consist mainly of the following two.
1) Access control to learning models (for users)
3.2 Learning model utilization framework This is a mechanism that allows users to select
To improve the accuracy of prediction and classifi- an appropriate learning model according to the usage
cation by machine learning technology, it is effective to purpose. Access control to learning models is imple-
create learning models according to specific purposes. mented through a set of APIs that appropriately select
In other words, the aim is to develop a learning model learning models and execute prediction and classifica-
specific to the purpose for each applicable design field tion processing, selecting for example a learning model
and further for each applicable theme within each field. for electrical design support in response to a user re-
Further, design support services utilizing ma- quest from an electrical CAD, or a learning model for
chine learning technology are expected to be utilized structural design support in response to a user request
from various usage environments including design from a structural system verification tool. The informa-
tools such as computer-aided design (CAD) tools and tion required for using the learning model is stored in a
computer-aided engineering (CAE) tools, and web database in the cloud, and by linking the APIs and the
browsers. For this reason, we are working to construct database, it becomes possible to use learning models
a learning model utilization framework that makes it according to the purpose.
easy to use multiple learning models from various 2) Registration/updating of learning models (for ser-
environments. vice providers)
Figure 6 shows the system configuration of the This is a mechanism for service providers using
learning model utilization framework. The utilization machine learning. The design support service provider
framework system is a web service on the Engineering registers learning models in a database for managing
Cloud. The Engineering Cloud is a cloud environment the learning models on the cloud using APIs for learn-
offered by Fujitsu for the engineering field.4) Websites ing model registration. As technology and design
and CAD/CAE tools that offer services utilizing machine processes applied to product design always progress,
learning technology access the utilization framework there is concern that deviation from actual product
on the cloud. design will occur if the learning model, once made, is
A web application programming interface (API) used as is.
using Web technology is used as the access API. The To prevent this, it is necessary to have a

User environment Engineering

RDB MONOZUKURI AI framework
Cloud
Learning model
Model library
management
Management AI engine
... Learning Learning Learning
server model model model
CAD tool
Registration/ Referencing/
Deletion Execution
Prediction
processing

CAE tool API API Machine learning control

[https] [https]

Browser, other applications

API [https] Learning model Service providers
registration/update

RDB: Relational database

Figure 6
Learning model utilization framework.

48 FUJITSU Sci. Tech. J., Vol. 53, No. 4 (July 2017)

 N. Nozaki et al.: Application of Artifi cial Intelligence Technology in Product Design

mechanism to periodically update the learning model We will also strengthen the following initiatives to
according to the development of new design technol- automate design and verification using CAD/CAE and
ogy and processes. In the utilization framework, the design assets.
processing necessary for such learning model updating The first initiative is to increase the number of
is also implemented with APIs. Service providers can cases in which machine learning technology is applied
update the learning model on the Engineering Cloud in the design field, and to accumulate know-how that
by using the APIs for updating learning models. so far has been dependent on people through the use
By constructing an operation environment for of information and communications technology (ICT).
services using machine learning as the common foun- We will address the various issues in product design
dation, we are aiming for a design environment that is using the solutions to problems encountered in product
beneficial to both users and developers of services. development, such as in the development of machine
learning methods and in determining the value of
4. Goals features. More such solutions will provide us with the
So far, we have described the application of means to create an ever greater number of machine
machine learning technology to the design field and learning models that can make available heretofore
its effects, and the MONOZUKURI AI framework as an people-based know-how.
application of AI to the FTCP integrated development The second initiative is to develop data standards
framework. We are currently working on the develop- that utilize design assets as learning data. Design
ment of FTCP 3.0 based on the concept of One Platform, data so far was not intended to be used for machine
and the MONOZUKURI AI framework is an important learning technology, but manpower is often required to
component of that. For details about One Platform, extract feature vectors and analyze correlations among
please refer to the paper “Cloud-based Development data. In this regard, we will establish methods to ef-
Platform for Next-Generation Monozukuri” also pub- ficiently associate various kinds of data accumulated in
lished in this issue.5) The configuration of FTCP 3.0 the past and to shape data.
is shown in Figure 7. The learning models on the Finally, on the topic of design that uses the lat-
MONOZUKURI AI framework will be constructed as de- est technology, of design scale and design constraints,
sign support services through the cloud in the future. measures need to be looked into by which to deal with

FTCP 3.0 Design support services

Automatic wiring

Layer number

Shape search

Service provision
estimation

...
DRC

Web API

Creation

Design tools Know-how MONOZUKURI AI framework

sharing system
Electrical Learning AI engine
development model
platform Design
data
Information Feature Learning AI engine
Structural collection extraction model
development
platform Know-how
Learning AI engine
model
Project
management Operation Learning
system history
AI engine
model

DRC: Design rule check

Figure 7
Configuration of FTCP 3.0.

FUJITSU Sci. Tech. J., Vol. 53, No. 4 (July 2017) 49

N. Nozaki et al.: Application of Artificial Intelligence Technology in Product Design

designs that deviate from existing designs. In the case Japanese).

of a technical breakthrough, sufficient consideration 4) M. Yasuda: Fujitsu’s Engineering Cloud. FUJITSU Sci.
Tech. J., Vol.48, No.4, pp.404−412 (2012).
such as how to utilize the design assets before and
http://www.fujitsu.com/global/documents/about/
after the breakthrough will be necessary. In the future, resources/publications/fstj/archives/vol48-4/
we will establish indicators that can quantitatively paper03.pdf
make such judgments. 5) E. Konno: Cloud-based Development Platform for Next-
Machine learning is good at “prediction/classifica- Generation Monozukuri. FUJITSU Sci. Tech. J., Vol. 53,
tion” and “clustering.” Rather than solving everything No. 4, pp.22–29 (2017).
http://www.fujitsu.com/global/documents/about/
with machine learning technology, we will work on
resources/publications/fstj/archives/vol53-4/
differentiated design and automated verification for paper03.pdf
“rule-based automation,” “automation by machine
learning,” and “hybrid automation combining rule-
based and machine learning automation” depending Naoyuki Nozaki
on the design issue at hand. Fujitsu Advanced Technologies Limited
Mr. Nozaki is currently engaged in the
creation of development platform environ-
5. Conclusion ments for product design.

In the IoT era, AI is expected to play a role in many

different situations. On the other hand, AI technology
at the present time is not necessarily universal, and
application areas will have to be identified within the
scope of answers obtained from probability and statis- Eiichi Konno
tics. By mechanically calculating the correlation of a Fujitsu Advanced Technologies Limited
Mr. Nozaki is currently engaged in the
plurality of feature vectors that people did not notice, creation of development platform environ-
AI can be used as a tool for deriving responses that ap- ments for product design.

proximate human experience and know-how. Using AI

to gather the experience and know-how that to date
has been considered the preserve of humans and
put it to work in ICT opens the way to the use of AI in
product design. The FTCP supports product design, and Mitsuru Sato
we intend to contribute to the further strengthening Fujitsu Advanced Technologies Limited
Mr. Nozaki is currently engaged in the
of Fujitsu�s, and Japan’s, Monozukuri by adding to the creation of development platform environ-
ments for product design.
FTCP AI options developed in collaboration with various
product design departments.

References
1) Fujitsu: Fujitsu Takes Systematic Approach to Artificial
Intelligence with “Human Centric AI Zinrai.” Makoto Sakairi
http://www.fujitsu.com/global/about/resources/news/ Fujitsu Advanced Technologies Limited
Mr. Sakairi is currently engaged in the cre-
press-releases/2015/1102-01.html
ation of structural platform environments.
2) A. Miyazawa et al.: Construction of Integrated Design
Development Environment and Its Deployment on
Cloud. FUJITSU Sci. Tech. J., Vol.48, No.4, pp.394–403
(2012).
http://www.fujitsu.com/global/documents/about/
resources/publications/fstj/archives/vol48-4/
paper02.pdf
3) T. Kaniwa et al.: The Challenge of Developing Design
Automation using Design Resources. The 2015 IEICE
Engineering Science/NOLTA Conference. 2015 (in

50 FUJITSU Sci. Tech. J., Vol. 53, No. 4 (July 2017)

 N. Nozaki et al.: Application of Artificial Intelligence Technology in Product Design

Toshiyuki Shibuya
Fujitsu Laboratories Ltd.
Mr. Shibuya is currently engaged in re-
search on product design automation.

Yuuji Kanazawa
Fujitsu Laboratories Ltd.
Mr. Shibuya is currently engaged in re-
search on product design automation.

Serban Georgescu
Fujitsu Laboratories of Europe Ltd.
Mr. Georgescu is currently engaged in the
development of applied machine learning
technology.

FUJITSU Sci. Tech. J., Vol. 53, No. 4 (July 2017) 51