IEEE Robotics & Automation Magazine - June 2023 - 24

status. The main drawback of such a method is that it cannot
determine the key factors affecting the equipment health
reflected by the sensory data since the sensory data can present
the current status only. To address such an issue, FDC should
deal with not only the sensory data, but also the other collected
data, including the recipe data, previous and postrecipe data,
historical data, etc., resulting in large amounts of data. With
such large amounts of collected detail data to be dealt with,
the fault classification and diagnosis might be done to be more
accurate as shown in [2], [8], [9], [10], and [11]. However, the
underlying infrastructures of the current FDC do not have the
capability to handle such complex data since the collected data
sets have large sizes and contain noise, missing values, and
heterogeneous structures (i.e., the data can be characterized
into the structured data and part-missing unstructured data).
Thus, a proper data handling structure needs to be taken into
account to improve the overall data handling capability. This
motivates us to present a Hadoop ecosystem-based infrastructure
for FDC.
The Hadoop system has been widely used in practical
applications, such as the real-world health care data analysis
in [16]. Also, as a main component in the Hadoop system,
MapReduce is widely used to deal with large-scale data sets,
such as the large-scale protein-protein interaction data analysis
in [17]. However, compared with the applications of the
Hadoop system in other fields, semiconductor manufacturing
systems are languid, as stated as follows:
■ As the most complex and automated manufacturing systems,
the data in semiconductor manufacturing systems
have heterogeneous structures and are more complex. It is
not easy to deal with this kind of data.
■ If the Hadoop system is applied to semiconductor manufacturing
systems, during the updating of a software system,
the production lines are not allowed to be shut down
since the stoppage of the production processes would
cause huge economical loss.
■ Semiconductor industries are relatively conservative. Their
data and wafer fabrication processes are highly confidential
such that there is no condition for researchers to do scientific
and application studies devoted into how the Hadoop
system is applied in semiconductor manufacturing.
To overcome the above-mentioned difficulties faced during
the implementation of the Hadoop system, by cooperating
with a semiconductor fab in China, we make the following
contributions in this work:
■ Restrictions of the current FDC system in semiconductor
manufacturing are presented and the benefits from the
implementation of a Hadoop ecosystem are well analyzed.
■ We present a Hadoop ecosystem rather than a single component
within the ecosystem to make the data processing
as a unified entity. Each component serves as its own
functionality to perform data storage, transfer, processing,
and maintenance.
■ A smooth migration path is proposed so as to decompose
the current FDC system into the business logic and data-processing
modules. By separating different working function24
IEEE ROBOTICS & AUTOMATION MAGAZINE JUNE 2023
alities for different modules, it is possible for each
component within the Hadoop ecosystem to serve for data
storage, processing, and maintenance. Therefore, the current
FDC system in fabs is decomposed, rather than newly developed.
In this way, the current FDC system for a fab can be
successfully migrated to a Hadoop ecosystem-based one
without shutting down a wafer fabrication line. Furthermore,
this migration method is also applicable to other software
systems, such as a manufacturing execution system (MES),
advanced planning system (APS), and the other main systems
integrated in the current CIM infrastructure.
The rest of the article is organized as follows. In the next
section we discuss the architecture of a Hadoop ecosystem.
Then, in the " Hadoop Ecosystem-Based FDC System " section,
we designed such a system. Also, a migration path from a
current FDC system to a Hadoop ecosystem-based one is presented
such that the migration can be smoothly done without
shutting down a wafer fabrication line. In the " Performance
Evaluation " section, experiments are carried out to verify
the performance of the system. Finally, we summarize in the
" Conclusion " section.
ARCHITECTURE OF A HADOOP ECOSYSTEM
RESTRICTIONS OF THE CURRENT FDC SYSTEMS
The international technology road map for semiconductors
defines the dimensions of the big data problem in terms of
" the five Vs, " referring to volume, velocity, variety, veracity,
and value [18]. In the past, single and dual-combined computer
integrated systems aimed to deal with one of the five
issues. However, with the development of IoT and AI technologies,
such five issues of the big data can be further investigated
by applying a Hadoop ecosystem so as to improve the
FDC system in semiconductor fabs.
In a modern fab, the amount of data that can be acquired
is continuously increasing with the implementation of IIoT
technologies. With the collected data, the fault detection and
diagnosis of the equipment is performed by an FDC system.
The more the data are collected, the more accurate the results
of the FDC system can output. However, the current FDC
framework is built on a single server based on a traditional client-server
architecture, which cannot deal with large amount
of data in a short time. Meanwhile, it has a restriction on handling
the unstructured data. Furthermore, the current FDC
system adopts a simple calculation for the data analysis such
that it fails to analyze the relationships among different features
of high-dimensional data. To address such an issue, great
efforts have been made on data analysis by AI-based methods,
such as machine learning [10] and artificial neural networks
[19]. Nevertheless, large data sets are required as the input for
an AI-based model, while the current FDC framework fails to
support dealing with large data sets. Furthermore, better data
collection brings better data quality. This can shorten the time
consumption for data processing since fewer preprocessing
operations of the collected data need to be done to improve the
data quality, prediction accuracy, etc.
IEEE Robotics & Automation Magazine - June 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2023
