IEEE Robotics & Automation Magazine - June 2023 - 30

TABLE 1. Hardware parameters.
CPU
Local database
FDC application
5220R @ 2.20 GHz 256G
5220R @ 2.20 GHz 256G
Hadoop ecosystem 5220R @ 2.20 GHz 256G
MEMORY NETWORK OPERATING SYSTEM SERVER TYPE
1G
1G
1G
CentOS 7.9.2009
CentOS 7.9.2009
CentOS 7.9.2009
Physical server
Physical server
Physical server
SERVER NUMBER
2
2
6
TABLE 2. The adaptability tests for bulk data loading.
Bulk data
loading
DATASET SIZE TIME USAGE DATA THROUGHPUT
30 GB
100 GB
1,000 GB
1 min 42 s 293.76 MB/s
5 min 31 s 322.44 MB/s
43 min 48 s 380.52 MB/s
of data sets are too large and they contain noise, missing values,
and heterogeneous structures. This results in that not all
detailed data can be collected and there are not enough data to
be used for the FDC system.
However, by the designed of the HE-FDCS, more detailed
TABLE 3. The adaptability tests for data analyzing and
sorting algorithms.
Means
DATASET SIZE TIME USAGE
30 GB
100 GB
Bayes
classification
Distributed
sorting
algorithm
(Terasort)
1,000 GB
30 GB
100 GB
1,000 GB
30 GB
100 GB
1,000 GB
3 min 20 s
6 min 56 s
73 min 49 s
45 s
2 min 21 s
23 min 49 s
1 min 49 s
5 min 30 s
56 min 53 s
DATA
THROUGHPUT
150 MB/s
240 MB/s
226 MB/s
666.67 MB/s
750 MB/s
699.79 MB/s
273.57 MB/s
302.78 MB/s
293.8 MB/s
test is conducted for bulk data loading and different data-analyzing
algorithms with three different data set sizes, including
30 GB, 100 GB, and 1,000 GB. In this way, the response
time and throughput speed at different data set size levels
(i.e., small, medium, and large size) can be observed. The test
results for bulk data loading are summarized in Table 2.
In practice, the bulk data loading is done every day at offpeak
hours. It follows from Table 2 that as the data set size
grows, the time needed for bulk data loading increases. For a
1,000-GB data set, it takes 43 min and 48 s for data loading.
However, it has to take 45 min on average to load a 30-GB
data set in the real fab. Also, it spends more than 2 h to load
such amounts of data under an extreme situation. By comparing
this with the current situation in the fab, HE-FDCS
is good enough to deal with a data set with 30 GB, 100 GB,
and 1,000 GB for bulk data loading within 2 min, 6 min, and
44 min, respectively.
With a real semiconductor fab as a test case, EAP systems
are used to collect the equipment health condition data. The
underlying infrastructures of the current FDC systems do not
have the capability to handle the collected data since the sizes
30 IEEE ROBOTICS & AUTOMATION MAGAZINE JUNE 2023
data of tools can be collected such that the FDC can perform
well based on machine learning methods. Thus, K-means and
Bayes classification algorithms are adopted to test the adaptability
of the designed HE-FDCS since they are the basic
methods for fault-cause analysis applied in fabs. The test
results are summarized in Table 3.
It follows from Table 3 that K-means and Bayes classification
can be performed within 1.5 h so as to obtain trained
models based on a 1,000-GB data set for fault diagnosis. Such
a training time is acceptable by the equipment engineers in
the fab. Notice that such a model training should be done
each day since the data sets are updated each day. Furthermore,
the EAP data frequency is 1,000 Hz, implying that the
data collection for a tool is done every .001 s. Such data are
used to map the health status of a tool. Once an abnormality is
detected by HE-FDCS, the analysis module at the FDC layer
could immediately respond such that the alert module starts
to work. At the same time, the fault diagnosis is immediately
done by the trained models obtained based on machine learning
algorithms at the FDC layer. From the above discussion,
the designed HE-FDCS can well support machine learning
methods for fault diagnosis.
In the meantime, the distributed sorting algorithm known
as TeraSort is adopted for measuring the sorting time duration
since it is often used in fabs. The test results are also summarized
in Table 3. From the test results, for a 30-GB data set, the
sorting time is 1 min and 49 s, which is slightly shorter than
that in practice. For 100-GB and 1,000-GB data sets, 5 min and
30 s and 56 min and 53 s are needed for data sorting, which are
also acceptable under the real manufacturing conditions.
PERFORMANCE TEST
TPC-DS [28] is a database query performance benchmark
protocol, consisting of a suite of business-oriented queries
and concurrent data modifications [29]. As widely used and
intended to offer fair and honest results for highly complex
data query templates on data query performance, TPC-DS
was applied to test the performance of HE-FDCS in this article,
including the CPU utilization, request success ratio,
memory utilization, and HE-FDCS response time.
Three scenarios with different data workloads as shown in
Table 4 are adopted for the performance test of HE-FDCS. The
IEEE Robotics & Automation Magazine - June 2023

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