Using DL Models in the Service Layer to Enhance the Fault Tolerance of IoT Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper addresses the problem of missing data in IoT networks. Moreover, the authors deploy a recurrent neural net (RNN) as well as a multi-layer perceptron (MLP) to predict this missing data.

General comments:

Note that the presented RNN cannot handle any kinds of missing data, and the MLP cannot do that either.
The authors must clarify how missing data is classified (Missing completely at random/Missing at random/Missing not at random), how to deal with the respective class (impute/interpolate/ignore)?
Following this approach, a large number of existing literature is missing. It is recommended to follow a systematic approach using PRISMA, to identify the missing records.

The authors are requested to define the requirements on the missing data for their
application purposes. This includes statistical properties and dynamics (i.e., the range of missing values with respect to the last available value in per cent).

For the experimental results, the MLP cannot handle any dynamics in the missing data and hence, it is unfair to use that for comparison. On the other hand, Dynamic Bayesian Networks are indeed capable of doing so (see for example the works in DOI 10.3390/s20113246 that is highly relevant but missing in the list of references. They show how such a network can handle missing sensor data as long as the underlying statistics does not change). For the sake of fairness, the authors are required to compare the performance of their RNN to that of such a dynamic Bayesian network, instead.   

Table 3 and Table 4 are irrelevant for the handling of missing data and hence,
can be safely removed.


Nevertheless, the study could potentially have interesting  on the behavior of reccurrent neural nets in the context of missing data.


Detailed comments:

Figure 1 is unreadable, and hence has to be replaced.

Author Response

Please see the attcahment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

In an IoT network, data from sensors is processed through gateways and controllers to service servers and then stored in the cloud. When devices fail, data gaps arise, rendering them unsuitable for analytics and predictions. This paper introduces recurrent neural networks (RNN) to predict missing data, comparing multi-layer perceptron (MLP) and long short-term memory (LSTM) models. The LSTM-based RNN outperforms other models with a prediction accuracy of 99.66%, making it a superior solution for data gap estimation.

The paper is interesting and deals with an important issue. Many sensor-based applications are emerging, but ensuring the flow of data without missing value is crucial.

The paper is written reasonably well but there is a need for improvement in various aspects. In this respect, this review would like to see the following changes:

+ In the first para of the “Introduction” section, the authors discuss the IoT layer. For the interest of readers, it would be better to explain it with a diagram.

+ Literature Review: This section needs improvement. There is very little discussion on some of the works, example [7], [10], [11], [13], [14]... The discussion should include what they have done, what are the merits and demerits of those works in comparison.

The authors should also include a few recently published works, such as see  (https://doi.org/10.1109/TR.2024.3416967; https://doi.org/10.1109/WiMob58348.2023.10187811). Those works estimate the next sensor data in the cycle with an intention to find reliability of the data, yielding greater fault tolerance. The aim of the current work is also to predict missing data. When data is missing, it can be replaced by estimated data. In this way, those works are related and should be discussed. Among the cited references in this [6]-[19], only two are published since 2020. Adding a few more recent relevant papers will strengthen the “Related Works” section.

Each sentence begins with a number, for example, “[6] H. Liu has expressed …” which is very strange. It should rather be “Liu [6] has expressed …”

+ Quality of Fig. 1 should be improved. In Fig.4, please mention the legends, what ate Ts, Fs, Ds,CHs …

+ In building models, the authors have used MLP and LSTM. Are there any particular reasons why these models were selected?

+ How will your model deal with ‘concept drift’? If the nature or range of the data changes, for example, the temperature data in some days change slowly and other days may change widely and several times depending on weather fluctuation. You should make some comments on this.

 

 

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Quality of the revised version has improved but there are a few (minor) issues.


1) Is the missing data under consideration "random (MAR)" or "completely at random (CMAR)"? In other words, can the missingness be predicted using other observed data?

The authors' reply "We have shown that sequence data is not dynamic and can
be handled using MLP, LSTM, using a regression Model." seems to indicate that the authors consider *static data* with *missing observations* i.e, the system is MAR. This information is highly important for understanding the rest of the paper and hence, must be added (right in the introduction).

2) Reference [20] does indeed consider MAR (see right after Equation (10) in their paper), and the range of the output signal is Real (not Boolean), in contrast to what is stated in the paragraph at line 141.  That is why this reference is highly relevant.

3) Reference numbering went wrong during compilation. This is true for [20] and [21] in line 141 and line 146, respectively.

Author Response

Please see the attachment

Author Response File: Author Response.pdf