Flush Airdata System on a Flying Wing Based on Machine Learning Algorithms

Round 1

Reviewer 1 Report

This paper presents the modeling results of a flush air data system (FADS) using neural networks. The results seem interesting, with good predictive capabilities, but the authors omitted many relevant aspects of the methodology. The section 5 ("Result and discussion") is limited to one paragraph and should be improved.

The method to define the number of pressure ports selected is only partially explained. Then, an ensemble of artificial neural networks (ANN) is used to predict angle of attack, sideslip and Mach number. Some of the neural networks are regarded as "low precision", whereas others are "high precision", but it is not clear to which ANN the results are related to. The ANNs also have different structures (number of neurons in the hidden layers), but the authors did not provide reasoning for this. The authors did not explain how the database for training was generated, nor which parameters are inputs for the ANNs. The manuscript is missing a number of key points, which are detailed below. For these reasons, at this point, the paper is not acceptable for publication.

Extensive editing of English language is also required. A number of typos and errors are listed below, but many others exist in the paper.

 

 

Comments regarding contents:

In Section 2, the authors mention that MIV was used to evaluate the relationship between variables in a neural network (line 152). Which neural network is this? (i.e. How many nodes in the hidden layer? What are the input variables? Pressures from 100 candidates pressure ports?)

In the pseudo-code shown (lines 157-166), what is "i"? Each independent variable? How many are they? In line 168, "N" is the number of tests. What do you mean by "test"? Is this the number of independent variables?

"The first ten port locations selected by two methods were applied to train the artificial neural networks respectively."
However, in section 5, line 447, the authors mention that 9 ports were used, as selected by using MIV. How many ports were used? 10 or 9?
Why 10 or 9 ports selected? What was the criteria used? Why not 5 or 15?

"The prediction errors were compared in Figure 2 and Figure 3"
What are the x-coordinates in these figures? Training sets for the ANN? How were they obtained? What kind of tests are these? Static/dynamic? Are high angles of attack included (i.e. stalls)?

The reason for doing so is that the flow separation characteristics are inconsistent at different altitudes
This is the first time "altitude" is mentioned. Is altitude an input variable? Was it varied in the training set? Or the Reynolds number was varied instead?

"To overcome this problem, the global optimization method is needed to aid the 220 training." There is no guarantee that these methods converge to the global minimal. Perhaps say "in order to minimize this issue, two optimization melhods were used..."

"Two optimization algorithms, genetic algorithm and Particle Swarm Optimization were applied in this paper to get a high precision artificial neural network."
I believe "high accuracy" should be used, instead of high precision.

In Figs 5 and 6, I don't understand why the "Train the network with L-M method" is necessary, if the optimal weights and biases were already obtained with GA or PSO. Was this a refinement?

The pseudo-codes for GA and PSO presented in section 3 are rather generic. Please define what properties were considered as genes, chromosomes, etc in your application.

"The artificial neural networks optimized by Particle Swarm Optimization (PSO) and genetic algorithm (GA) used 20 neurons in the hidden layer for the angle of attack and the angle of sideslip, and 10 neurons for Mach number; while the number of hidden layer neurons in the unoptimized neural network is 50 for the angle of attack and the angle of sideslip, and 20 for Mach."
Why these specific numbers? What are these optimized and unopmitized ANNs? I assume they are related to the ANNs shown in Figure 4, but the authors have called these "low precision" and "high precision", in lines 209-216. (Again, I believe the term should be high/low accuracy).

"All the optimized artificial neural networks have less average error than original network,"
What is the original network? Is this the one also called "unoptimized" before?

It is really difficult to see any differences between the six optimized ANNs in Figures 7 and 8. Perhaps show an histogram of the error, since the x-coordinate (order of experiments) is not relevant.

The results in Figures 7 and 8 are for the ensemble of ANNs? Or for one ANN in particular? This is not clear.

"thus the PSO method was used to train all the rest of the artificial neural networks"
So, the results in figures 7,8,9 are for a single ANN? Which one?




Comments regarding format:

The subsections in Sections 3 and 4 are numbered incorrectly.

throughout the manuscript:
instead of "angle of slide", the usual term is "angle of sideslip".

Many typos and grammar errors are indicated below (but not all of them):

in addition, the number of the transducer
in addition, the number of transducers

thus the normal method for FADS system is difficult to maintain the accuracy.
making it difficult to maintain accuracy in the normal method for FADS systems

unmanned aircraft vehicle (UAV)
unmanned aerial vehicle (UAV)

degraded flying handling qualities
please decide "flying qualities" or "handling qualities". These terms mean different aspects in flight testing.

the locations of the pressure ports have some restrains
the locations of the pressure ports have some constraints

recently, Jiang et al. [8] combine
Recently, Jiang et al. [8] combined

though the mothed can
although the model can

"measurement location in FADS system far less strict than it was. However, this method still has some restrains"
measurement location in FADS system far less restrictive than before. However, this method still has some constraints"

all the airdata (including angle of attack, sideslip angle, Mach number, etc.) is taken as output, and the neural network is used to establish some mapping relationship between the input and output vectors (the relationship between the two is very complex and highly nonlinear).
the airdata parameters of interest (e.g. angle of attack, sideslip angle, Mach number) are taken as output, and the neural network is used to establish a mapping relationship between the input and output vectors. Typically, the input-output relationship is very complex and highly nonlinear.

has a high real-time quality
has good real-time performance

small unmanned aircraft vehicles (sUAV)
small unmanned aerial vehicles (sUAV)

replace the fault port with adjacent ports
replace the faulty port with adjacent ports

which restrain the locations of the transducers
which limit/constrain the locations of the transducers

thus it is not allowed to install the pressure
thus it may not be possible to install the pressure

with high efficiency and precision.
perhaps "accuracy" instead of precision

6. where N represents the number of tests.
where N represents the number of tests.

can be seen form the figures
can be seen from the figures

The pressures form the sensors
The pressures from the sensors

The pressures form all the ports
The pressures from all the ports
(Several other instances of "form" used instead of "from" must be revised)

Therefore, the fault sensor needs
Therefore, the faulty sensor needs

Four different machine learning methods by identifying the correct pressure combinations form the error ones was tested for the error diagnosis of flush airdata system.
(this sentence makes no sense, please revise)

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The work presented in your paper is highly interesting and after you address all my comments (please see the attached file) it will be ready for publication.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Thank you for addressing all the comments. I believe the manuscript is acceptable for publication in the current form.

Reviewer 2 Report

Thank you for addressing all my comments. I recommend the manuscript be published.