IEEE Computational Intelligence Magazine - February 2022 - 105

recorded failure (correct disks), it can be
assumed that all the windows are correct.
Nevertheless, in a failed disk, it can be confidently
considered that when it is initially
put into operation it works correctly for a
while. Then, at a certain moment, it begins
to show signs of malfunction until it suffers
a breakdown and stops working. Consequently,
it can be assumed that the first
window (when the disk is put into operation)
is a correct window (class 0) and that
the last window, just before the breakdown,
shows signs of malfunction and it is a malfunction
window (class 1) (see the problems
highlighted in Figure 1). Since the
moment when the signs of malfunction
begin to appear is unknown, the label of
the rest of the windows is unknown. However,
in order to learn a classifier which is
able to detect windows with signs of malfunction,
it is mandatory to have examples
of these kind of windows. Therefore, for
failed disks, a labeling process is necessary
to identify those windows.
A double-round learning methodology
is proposed to solve this problem. It
is thus named because the training process
is done in two stages (See Figure 2),
which are explained below:
1.1 Generate a training dataset with
one window from each hard
drive: In this first step, the purpose is
to generate a dataset in which the correct
and malfunction windows are
clear cases of correct operation and
malfunction. It is assumed that this
occurs when the correct windows are
selected from the beginning of the
operation of the disk and when the
malfunction windows are selected
from just before the failure occurs.
Therefore, for correct hard drives, the
first window of observations is selected
and, in the case of failed hard drives,
the last window of observations is chosen.
Consequently, the generated training
dataset comprises only one
window of each disk available and will
be referred to as One-window dataset.
1.2 Train the first classifier: The first
classifier is trained with the dataset
created in the evious step, the Onewindow
dataset, (the details about
the classifiers used are explained in
Section II-C).
2.1 Apply the classifier 1 to all the
windows of the failed disks and
create a new training dataset:
The classifier trained in the previous
step is applied to all the windows of
the failed disks of the training dataset.
Contrary to other proposals in the literature,
which preset a common lead
time value for all disks, this is done to
1.1 Generate a Training
Dataset With One Window
From Each Hard Drive
W
W
W
W
W
attempt to automatically detect when
the disks begin to malfunction and
label the windows of the failed disks.
Specifically, the purpose of this step is
to identify representative windows
that show signs of malfunction and
are not at the end of the MTS.
For the failed disks in which the
last window is not correctly classified
1.2 Train
the First Classifier
One-Window Dataset
1
1
Classifier 1
2.1 Apply Classifier 1 to All
the Windows of the Failed
Disks and Create a New
Training Dataset
Failed Hard Drives
Disk 1
Disk 2
Disk 3
Classifier 1
Classifier 1
1
Classifier 1
Correct Hard Drives
FIGURE 2 Diagram representing the steps of the proposed double-round learning methodology.
1
1
1
1
1
1
1
1
1
1
2.2 Train the Second
Classifier With the
Labeled-Window Dataset
Labeled-Window
Dataset
Classifier 2
FEBRUARY 2022 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 105

IEEE Computational Intelligence Magazine - February 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2022