Supporting Adaptive Coding and Modulation Techniques for Satellite Radio Channel

(1)

Supporting Adaptive Coding and Modulation Techniques for Satellite Radio Channel

Árpád László Makara

Department of Broadband Infocommunications and Electromagnetic Theory,

Faculty of Electrical Engineering and Informatics, Budapest University of Technology and Economics,

M ˝uegyetem rkp. 3., H-1111 Budapest, Hungary.

makara.arpadlaszlo@edu.bme.hu

László Csurgai-Horváth

Department of Broadband Infocommunications and Electromagnetic Theory,

Faculty of Electrical Engineering and Informatics, Budapest University of Technology and Economics,

M ˝uegyetem rkp. 3., H-1111 Budapest, Hungary.

csurgai-horvath.laszlo@vik.bme.hu

Abstract—For satellite-Earth communications, higher fre- quency bands, especially with millimeter wavelength frequen- cies are very sensitive to precipitation and other atmospheric changes. The time variation of these effects is extremely variable, eventually increasing the current attenuation of the channel. Taken together, these phenomena may significantly reduce the momentary capacity of the channel. The primary goal is to transfer as much data as possible correctly on a given link, which means that the optimal protocol should be used at each time step.

The current research is built around two main pillars: on the one hand, it takes stock of existing methods for ACM problems in satellite communications, and; on the other hand, we investigated how to implement state estimation for real interconnections. This was based on actual measurements from the Alphasat satellite. The ModCod settings for DVB-2 transmission are estimated from the beacon signal transmitted by the spacecraft. For the implementation, we used a deep neural network with memory, the results of which can be used as a basis for future tests to solve similar problems.

Index Terms—satellite links, time series prediction, fading, ACM, AI, DL

I. INTRODUCTION

Satellite-to-earth links differ from other wireless communication methods in that the channel is passing through the Earth’s atmosphere; in addition to all, it is a direct link. In terms of electromagnetic signal propagation, two atmospheric layers are significantly affecting the propagation: the thermosphere and the inosphere [1]. While the former is the Earth’s lower atmosphere [2], where weather phenomena take place, the latter is the higher region of the atmosphere, where gases are ionised by the Sun [3].

Electromagnetic waves (EM waves) of different wavelengths are affected differently. For practical use, 1GHz is considered the limit above which ionospheric effects will be negligible [1]; which is why they are mainly used for satellite communications at higher frequencies, such as millimetre waves (mmWave). The methods presented are primarily designed for mmWave, where only the critical effects of the thermosphere need to be addressed.

The gases in the atmosphere absorb EM waves to different degrees, depending on the frequency [1], [4]. Of particular interest in mmWave is water and its different forms, which attenuate differently depending on their form, intensity (if it is rain), etc.. In summary, mmWave waves are sensitive to changes in the weather, which is manifested by changes in attenuation. As the attenuation varies, the physical communication channel, with a constant radiated power, does not have a constant capacity [5].

There are two options for communication: somehow choose a constant transmission settings, or adaptively change the coding procedures or the modulation used (ACM). We have at least three ways to categorise the possible solutions to implements ACM [6]. One of the classical solutions is to estimate a channel characteristic, typically SNR, from the measured signal levels, on the basis of which the change is made. Such methods, with appropriately rapid sampling, may be sufficient for channel variations. The former can be seen as a further development of procedures that already estimate these parameters for the future. Obviously, estimation is generally a more difficult operation to perform, but it also allows the system to react in advance to future changes [7]–[11]. A more complex and not yet fully researched area is when, instead of all these procedures, we predict the corresponding ACM settings as a kind of state estimation (where state represents a pair of encoding and modulation settings) [6]. Solutions of this kind are based on some kind of artificial intelligence or machine learning. Thus, in order to implement them, a large amount of data is needed to perform an acceptable training process. Compared to the first two solutions, it has the advantage of combining two logical steps: estimation of the quantity and prediction based on the estimated quantity. Summing up the advantages and disadvantages, if the resulting algorithm is given enough resources to run in real time, it could theoretically be faster and more accurate than previous solutions.

The structure of the paper is the following: In sec. II.

the main features of the radio channel are presented.

(2)

Afterwards, in sec. III., there is a short summary of the most important existing procedures. In sec. IV. we detail our own results. Finally, sec. V is the conclusion.

II. ACMINSATELLITERADIOCHANNEL

In order to review the possible solutions, it is necessary to take a deeper look at the main factors influencing connec- tivity and the ACM processes that have been implemented in the past.

A. Satellite Radio Channel Main Parameters

In the millimeter wavebands, tropospheric effects are the most dominant in terms of attenuation. However, other phenomena are also important to note, the key factors affecting the quality of the connection is [4], [5], [12], [13]:

1) Rain fade (to be highlighted for its prominent role), 2) Rain, water, cloud (in summary, these are called hyd-

rometeors) absorption, depolarization and scattering, 3) Absorption by atmospheric gases (other than hydro-

meteors),

4) Inclined orbit satellite,

5) Antenna pointing errors (direct link, long distance so need highly directional antennas),

6) Noise,

7) Scintillation (impact of the ionosphere), multipath effect (interference etc.),

8) other more complex parameters [5].

The impact of tropospheric phenomena, i.e. weather, can be measured to a moderate extent. In most cases, the connection is slant. This means that the line of connection forms an acute angle with the surface of the earth (at the point of the ground station (GS)). Therefore, what is important is not only the weather that can be measured locally, but also the weather that can be experienced further away from the station.

B. Basics of ACM

For each satellite link, a link budget calculation is required for every GS [14], [15]. This gives the main parameters of the links on which the communication is to be implemented. A real example is shown in Table I.

For each link, we can determine a minimal noise value per symbol energy

³Eo No|min

´

along the descriptive statistics.

This value will actually tell us what the typical nature of the channel is, given the specified accuracy and availability.

That is, at this value, communication can usually be suc- cessful with the right modulation and encoding [13]. Fig. 1.

shows a schematic of this situation, with the gap marked in blue.

However, there may be situations where the minimum value is exceeded. In this case, to increase the amount of data transferred, you need to change the settings. This procedure is essentially the ACM. In other words, the aim is that the somehow chosen limit should be as close as possible to the instantaneous capacity of the channel. As illustrated in Fig. 1., the role of the ACM is to make the

Fig. 1: Theoretical benefits of adaptive coding and modulation techniques, schematic diagram. [12].

TABLE I: Link budget calculation for GS in Budapest with Alphasat [15].This GS provided the data for the results presented in this article.

Parameter Value Unit

Frequency 38.1 GHz

Guaranteed EIRP 38 dBW

Earth-satellite slant path 38 300 km

Free-space attenuation 215.72 dB

Atmospheric loss (estimated) 0.5 dB Ionospheric loss (estimated) 0.8 dB

Contour loss (estimated) 2 db

Reciever antenna diameter 1.8 m

Reciever antenna efficiency (estimated) 20 %

Receiver antenna gain 50.13 dBi

Receiver bandwidth 1.25 MHz

LNA noise figure 3.0 dB

Received signal power -98.89 dBm

Received noise power -109.03 dBm

C/N_o 71.11 dB

Es/No (maximal value) 10.14 dB

red line (established boundary) reach the edge of the green range. A possible choice of modulation and coding pairs, mentioned earlier for GS, is given in Table II.

TABLE II: DVB-S2 ModCod settings for a given bit error rate (PER = 10⁻⁵) [15], [16]. Among the possible settings, only those that can be used for communication between Budapest GS and Alphasat are listed in this table.

ModCod Required Es/No, Receiver Fade Spectral Efficienty

dB Margin, dB bits/Hz

QPSK 1/2 1.46 8.68 0.63

QPSK 3/5 2.86 7.28 1.00

QPSK 2/3 3.66 6.48 1.11

QPSK 3/4 4.36 6.78 1.25

QPSK 4/5 5.16 4.98 1.33

QPSK 5/6 5.56 4.58 1.39

QPSK 8/9 6.66 3.48 1.48

8PSK 2/3 7.16 2.98 1.67

8PSK 3/4 8.46 1.68 1.68

16APSK 2/3 9.66 0.48 2.50

(3)

III. USEDACM METHOD FORSATELLITECOMMUNICATIONS

It is clear from the above that the use of ACM is essential for effective communication. At the same time, an estimate of the expected condition of the channel is needed to make this happen. For this, we have no other information than the measured signal level. In some cases, this may be supplemented by a range of weather data, but there can be a lot of uncertainty. Ultimately, any estimate is based on past values for the future, but there is a difference in exactly which value is being estimated.

The simplest procedures calculate a kind of expected value from the SNR values already established [9], [11], [15].

A. Machine Learning-based Solutions

Except for the cases where the ACM procedure is chosen without using the data measured on the channel, we always use some kind of machine learning (ML) [17] solution. In many cases, these processes are called artificial intelligence- based [6]. The two concepts are not the same, although they are related to each other. Artificial intelligence is a field of computer science that makes a computer system that can mimic human intelligence [6], [18]. Machine learning is a subfield that enables the machine to learn from past data without being explicitly programmed [6], [18].

It is true for any machine learning solution that it is only good for what it was taught to do. For this reason, if the statistics of the parameters from the operating point change, it is necessary to adapt the change (of course, other unexpected effects and experiences may occur, but the primary purpose of any such algorithm is). Therefore, it is an important parameter for wevery ML solutions, how adaptable and maintainable it is.

Solution A

Solution B

ACM

AI to

estimate AI to decide

ACM

AI to estimate and decide

Measurements

Communication system

Fig. 2: Possible implementations for ACM using ML or AI in generally. [6].

The ACM procedure has to solve two tasks in real time:

an estimation problem and a decision problem (Fig. 2.). The estimation task is usually the harder one, it is necessary to estimate a channel characteristic quantity from some input data set. Such parameters could be SNR, receive signal level, Channel State Information (CSI), etc.

The decision task is to choose the best of the possible values based on the estimated parameters. The best choice

is to choose the setting with the highest throughput for a given bit error rate.

B. Experience Available for mmWave Satellite Connections with ACM

The connection used for testing is Q-band, with a nomi- nal frequency of 38.1 GHz [15]. The satellite used by the GS, is the Alphasat [19], for which an SNR estimation procedure has been carried out [10]. This is an online random regression forest (ORRF [20], an online algorithm that builds a random forest [21], [22], which can predict the SNR. A part of the result of the procedure is shown in Fig. 3. Based on the results available, we actually get a sort of moving average of the SNR.

In particular a node only splits if (i) a minimum number of samples has already passed the node (which ensures statistical significance), (ii) the depth of the tree has not exceeded the predetermined maximum model complexity (which limits the size model), and (iii) the minimal information gain required by a split is reached (avoids early growing). The objective function for the information gain is the sum of all target value differences between the observed true value and the predictions of the node for all parent and left/right child nodes.

A specific challenge introduced by online processing in our task is the unknown/changing range of feature dimensions. Extremely random forest approaches select random dimensions and random thresholds of the samples for decision candidates (weak learners). In the offline case, feature ranges can be trivially calculated from the whole dataset, but in the online case, this is usually not possible. It is reasonable to carefully select and adapt the feature dimension’s boundaries, to avoid useless calculations and a waste of memory.

We allow changing the feature dimension boundaries during the growth of the tree, starting with very conservative feature ranges and extending them dynamically to the highest/lowest values seen so far.

In the presented activity, we use a C++ implementation of the ORRF. The number of trees in the forest has been set to 20 with a maximum depth of 10. An example for the output of the ORRFs short time prediction done on a sample QV-Band SNR trace is shown in Fig. 3 assuming a dead time of 1s. As illustrated, the ORRF is following the real SNR time series quite closely, although the amplitude of the variations is lower. Compared to a baseline of assuming the current sample to be an estimated for the sample after dead time, the absolute prediction error is 45% lower.

Note that the ORRF just provides a prediction of the SNR.

Determining the modcod directly on this prediction still leads potentially to a too high FER, such that we added a fixed offset to the predicted SNR value as basis for modcod selection. Due to the online learning approach the algorithm is able to adapt to setups with other conditions after some minutes of operation.

Fig. 3. Example of true and predicted SNR time series using ORRF and with an ACM dead time of 1s.

IV. DATA AND SIMULATION APPROACH

The performance evaluation presented in this paper is based on power values of the Q-band beacon of Alphasat received by the 3m Q/V-band ground station in Graz. The power values are measured at a sampling rate of 10 samples/s.

Postprocessing of the beacon data included removal of outliers, patterns and ground station issues like antenna pointing problems. The resulting time series consists of 640 million samples, which corresponds to a time duration of about 2 years.

In order to model the SNR variations considering up and downlink, we scaled the Q-band time series to V-band using instantaneous frequency scaling from ITU-R recommendation P.618-11 [9], and add a time shifted copy to the Q-band data.

The resulting time series was scaled to the dynamic range of the considered modcods. The maximum SNR was fixed at 15dB, whereas the minimum was set to -2.5dB. Samples that went below the minimum SNR where removed from the time series and not considered further in the evaluation. The data has been split into 66 chunks, i.e., each chunk contains about 5.5 days of data. Each of the chunks has been split into training (50%), validation (20%) and test (30%) data.

We set up a simulation framework, which is able to evaluate the whole data series for different algorithms and parameterizations. The decision for switching the modcod is done for every sampling time instant k, and is based on the time series up to the sample 𝑠𝑘. Then the SNR threshold 𝑆𝑁𝑅𝑇 of the modcod is compared with the effective SNR value 𝑇_𝐴𝐶𝑀 later.

The calculation of the FER is based on a model reflecting the error behaviour of the channel codecs used in DVB- S2/S2x. For simplicity reasons, we assumed similar error behaviour of all modcods with respect to their SNR threshold, such that we could define a function 𝐹𝐸𝑅𝑐 normalized to 0dB for a FER of 10^-5 as shown in Fig. 4. The instantaneous FER at sample k, i.e. 𝐹𝐸𝑅𝑘, is then calculated by

 𝐹𝐸𝑅𝑘= 𝐹𝐸𝑅𝑐(𝑠𝑘+𝑇_𝐴𝐶𝑀⁄𝑇_𝑠− 𝑆𝑁𝑅𝑇) 

if 𝑆𝑁𝑅𝑇 is lower than the effective SNR value 𝑠𝑘+𝑇_𝐴𝐶𝑀⁄𝑇_𝑠. Otherwise 𝐹𝐸𝑅𝑘 is set to 0. Averaging over all 𝐹𝐸𝑅𝑘 gives the total FER, i.e. with 𝑁 being the total number of samples

 𝐹𝐸𝑅 =¹_𝑁∑^𝑁𝑘=1𝐹𝐸𝑅𝑘. 

For evaluation of the spectral efficiency, we use DVB- S2/S2x modcods from QPSK-1/4 up to 32-APSK-8/9 corresponding to spectral efficiencies from 0.4 to 4.44 bit/s/Hz. The resulting average spectral efficiency 𝑆𝐸𝑎𝑣𝑔 is calculated by

 𝑆𝐸𝑎𝑣𝑔=_𝑁¹∑ 𝑛𝑚 _𝑚𝑆𝐸𝑚, 

where 𝑆𝐸𝑚 is the spectral efficiency of modcod 𝑚 and 𝑛𝑚

are the number of samples, where modcod 𝑚 was active.

Authorized licensed use limited to: Dedan Kimathi University of Technology. Downloaded on May 18,2021 at 02:56:20 UTC from IEEE Xplore. Restrictions apply.

Fig. 3: ORRF-based SNR prediction for the Alphasat satellite [10].

Summarizing this guideline, the problem can be solved and the SNR can be estimated, on the basis of which the ModCod inputs can be set. There are a few open questions:

• What is the computational demand of the procedure for the desired accuracy?

• Random forest type solutions are difficult to maintain over time. Thus, it is questionable how easily they can be retrained if they are to be used in a new communication channel.

• How easily can this result be achieved by other methods?

In fact, one of the main motivations for the current research is to answer this last question. At the same time, instead of trying to solve a regression-type problem, we are thinking in terms of classification-based solutions. The rationale behind this is that for a real link, the number of possible modulation code pairs is finite. In other words, for us this can be seen as a classification problem, which in many cases is easier to solve than a regression problem (Fig. 2. Solution B).

Two other research results using neural networks (NN) are worth mentioning. Neural networks and deep neural networks (DNN) are structures whose neurons are inspired by the neural networks of living organisms [23], [24]. In the general case an elementary neurons can be described:

(4)

a=f¡P⃗W⃗ +b¢

, (1)

where a is the output of the neuron, f is the activation function,⃗P is the input vector,W⃗ is the weight vectorbis the bias. Of course, we can arrange individual elementary neurons into layers, feed back their outputs and thus create a variety of structures.

One interesting result of this topic was the ability to estimate the attenuation (in Ka band) using neural networks [25]. From all this, it is clear that such problems are not only theoretically solvable, but that the necessary computational capacity to apply DNN is already available.

The other DNN application to be mentioned is the one where the required parameters and feedback were investigated [8]. This brings us to two important pieces of information:

1) Knowledge of weather data is ultimately not necessary, but it certainly improves the result.

2) Using past values greatly improves the results.

In general, it is true that the more input data we use, the more complex the procedure.

IV. OURACM PREDICTIONMETHODS

The main goal of our research is how to implement estimation and possibly decision as a prediction problem.

Solutions of this type are expected to result in simpler algorithms with less computational effort than regression- based ones. We have tested two solutions:

• Based on fading prediction.

• ModCod status predictor.

Of course, if we change the encoding or modulation, the other device involved in the communication must be informed. However, the how and when of this is not discussed in the current phase of research. Nevertheless, it can be said that, with the same precision, the faster is the better.

A. Fading Prediction

The idea is simple: the most significant influencing factor, fading, is categorised. The algorithm estimates the expected fading, and the ModCod settings can be adjusted accordingly [26]. For this purpose, supervised machine learning should be used, which requires the input data to be labelled [17].

However, there is currently no procedure or algorithm to do this, so the process is done manually. Therefore, the primary tests were carried out for a simplified case, binary classification [26]. In the binary case of fading labeling, we considered a local, drastic signal level drop as a fadig. The signal level at the beginning of the mark and the signal level at the end of the mark must match, so this is the primary way to mark the edges of the label.

The substantive structure of the procedure (the parameter estimation) is shown in Fig. 4. According to the previous statements, only the signal level to be taken, and

Fig. 4: Binary fading prediction structure, part [26].

its function transformers, were calibrated; with a DNN that has memory. Memory is implemented by the so-called long short-term memory layer (LSTM) in the neural network [27].

This layer has internal states that enable it to explore and learn relationships in time series.

The input signal itself is more interesting. On the given GS, two signals are available [15]:

1) 38.1 GHz, DVB-S2 signal.

2) 39.4 GHz, Q band beacon signal.

The two signals behave remarkably similarly, so only the latter is saved in terms of signal level [15]. Hence, all our ACM procedures are based on estimating and deciding on the basis of this beacon signal, which then modifies the DVB-S2 settings on another frequency. The signal distribution for satellite links has a log-normal distribution– [28], which is also true in this case.

Based on the tests carried out, the accuracy of the DNN exceeds 93.3% [15]. This was achieved using one-minute sampling data. Nevertheless, fading projection needs to be improved, as it is not sufficiently accurate in many cases [15]. The method used depends on manually labelled data, which may have introduced a large degree of error into the training process. Thus, we are currently working on how to produce a more consistent algorithm-driven fading labeling procedure.

B. ModCod status predictor

The main idea during the creation of the procedure is to combine the previous statements for ACM. On the one hand, apply classification, not regression. On the other hand, combine the estimation and decision phases to logically speed up the process. Hopefully, the resulting procedure will be somewhat faster in runtime in practical tests.

The structure to use is basically the same as before, as illustrated in Fig. 4. We are changing two things along the experience so far. First of all, we are feeding back as input the ModCod setting that was used before. As it is available, it is particularly useful for training, thereby improving accuracy. The other change is that an extra layer of LSTM has been added to the DNN structure.

(5)

Due to the fact that this is also the form of supervised machine learning used, it is necessary to label the expected outputs during the training. For this purpose, we have the estimated SNR from the saved data [15].

The SNR estimation is based on a blind clock recovery using the Gardner timing error detection algorithm. The frame synchronization mechanism is based on the physical layer header (PLHEADER) that can be used also to decode the actual ModCod of the received frame. The decoded symbols are fed to the SNR estimator, which works in a non data-aided mode using the second-order and fourth- order moments methods. The instantaneous value of the SNR controls the ACM operation (Fig. 5); furthermore, its value is continuously logged together with the Graz pilot signal strength. The SNR estimation time is approximately 5 ms [15]. Because of our limited computational capacity, we worked with a sampling rate of about 10 ms for the rest of the time.

Fig. 5: DVB-S2 signal level and received ModCod settings [15].

Once the SNR value has been assigned to the time window, we select the ModCod setting that will transfer the most data from the available ModCod settings. DNN is required to determine this value based on the input data presented.

TABLE III: Test label method 1. In this procedure, we do not actually use the DVB-S2 signal, but label the ModCod beacons based on the beacon signal.

Used signal: min max Categroy

Q band beacon signal to label the output

non - 100 dB - 1 (no signal)

-100 dB - 90 dB 0

- 90 dB - 70 dB 1

- 70 dB - 55 dB 2

- 55 dB non 3

The tests so far have been done with simplified options:

only 4 ModCod settings have been applied to the saved data (the procedure is shown in Table III). The two edge settings (the most robust and the most sensitive), the other two settings were two imaginary ones with a combination of internal options. Therefore, there is no significant difference in the frequency of occurrence between the occurrence

of each option. It is important to highlight that for this approach, the most robust one is the best choice for the special case where there is no ModCod setting that meets the minimum conditions. It is still useful to treat them as a separate category, since the result will be the same from a control point of view, but not for the same reason.

TABLE IV: Test label method 2. The boundaries drawn are arbitrary, the aim was limited to simulating the expected category distribution.

Used signal: min max Categroy

DVB-S2 signal to label the output

non - 0.5 dB - 1 (no signal)

-0.5 dB 0 dB 0

0 dB 1 dB 1

1 dB 2.5 dB 2

2.5 dB 3.1 dB 3

3.1 dB 4 dB 4

4 dB 5 dB 5

5 dB non 6

The primary consideration for the second test labeling was to make the probabilities of occurrence of each category as realistic as possible (the procedure is shown in Table IV). Since the distribution of the signal used is lognormal, if you want to use as many categories as possible, you will have classes with very small sample sizes.

This means that the distribution of categories will be highly imbalanced, making training difficult.

Nevertheless, the error is not symmetric in terms of utilisation, i.e. the error of not using maximum channel capacity is not the same as the error of not transmitting data. Hence, we used a cross-entropy loss function [29] with an Adam optimizer [30] for the training.

The primary results are acceptable, with 84% accuracy for the case tested. However, as we increased the number of possible categories (to 8) and retrained the mesh, accuracy dropped to 27%. Accuracy of more than 60% can be achieved by rescaling the LSTM layer.

Examining the distribution of categories, we find that the more possible ModCod categories we use, the more uneven their distribution becomes and the worse the results. In our view, the primary problem lies in the applied error function. In any case, we are currently working on a suitable, sufficiently universal structure (which does not depend on the number of categories used to such a drastic extent).

V. CONCLUSION

With growing user demand, it is increasingly essential to maximise the use of the available capacity over a radio channel. There are several ways to solve the problem using machine learning. However, their reusability is not the same. The possible ACM settings used so far make it difficult to teach, which is one of the main problems of the method.

(6)

The solutions presented by us works satisfactory. At the same time, both procedures, albeit for different reasons, still need further development.

ACKNOWLEDGMENT

This research was supported by the Hungarian Ministry of Culture and Innovation and the National Talent Pro- gramme.

REFERENCES

[1] “Recommendation ITU-R P.676-12 (08/2019) - Attenuation by atmospheric gases and related effects - P Series -Radiowave propagation,”

Aug. 2019.

[2] R. Acharya, “Tropospheric propagation and impairments,” inSatellite Signal Propagation, Impairments and Mitigation. Elsevier, 2017, pp. 159–194. [Online]. Available: https://linkinghub.elsevier.com/

retrieve/pii/B9780128097328000065

[3] Acharya, Rajat, “Ionosphere and ionospheric impairments,” in Satellite Signal Propagation, Impairments and Mitigation. Elsevier, 2017, pp. 87–125. [Online]. Available: https://linkinghub.elsevier.

com/retrieve/pii/B9780128097328000041

[4] D. M. Pozar,Microwave Engineering, 3rd ed. John Wiley & Sons, Inc., N.J., 2005.

[5] “Recommendation ITU-R P.618-13 : Propagation data and prediction methods required for the design of Earth-space telecommunication systems.” [Online]. Available: https://www.itu.int/rec/R-REC-P.618/en [6] Á. L. Makara and L. Csurgai-Horváth, “Application of Artificial Intelligence in Satellite Communications,” in Selected papers of the 7th International Conference on Research, Technology and Education of Space (H-SPACE2022), 2022, pp. 1–5.

[Online]. Available: https://space.bme.hu/wp-content/uploads/2022/

09/Proceedings_Papers_HSPACE-2022.pdf

[7] Y. Yu and L. Zhang, “Adaptive Modulation Scheme for Satellite Communication Channel Based on RLNN,” Journal of Physics:

Conference Series, vol. 1856, no. 1, p. 012053, Apr. 2021. [Online].

Available: https://iopscience.iop.org/article/10.1088/1742-6596/1856/

1/012053

[8] L. Bai, C.-X. Wang, Q. Xu, S. Ventouras, and G. Goussetis, “Prediction of Channel Excess Attenuation for Satellite Communication Systems at Q -Band Using Artificial Neural Network,” IEEE Antennas and Wireless Propagation Letters, vol. 18, no. 11, pp. 2235–2239, Nov. 2019.

[Online]. Available: https://ieeexplore.ieee.org/document/8786170/

[9] H. Bischl, H. Brandt, T. de Cola, R. De Gaudenzi, E. Eberlein, N. Girault, E. Alberty, S. Lipp, R. Rinaldo, B. Rislow, J. A. Skard, J. Tousch, and G. Ulbricht, “Adaptive coding and modulation for satellite broadband networks: From theory to practice: ACM FOR SATELLITE BROADBAND NETWORKS,”International Journal of Satellite Communications and Networking, vol. 28, no. 2, pp. 59–111, Mar. 2010. [Online]. Available: https://onlinelibrary.wiley.com/doi/

10.1002/sat.932

[10] J. Ebert, W. Bailer, J. Flavio, K. Plimon, and M. Winter, “A method for acm on q/v-band satellite links based on artificial intelligence,”

in2020 10th Advanced Satellite Multimedia Systems Conference and the 16th Signal Processing for Space Communications Workshop (ASMS/SPSC), 2020, pp. 1–5.

[11] S. Cioni, R. De Gaudenzi, and R. Rinaldo, “Channel estimation and physical layer adaptation techniques for satellite networks exploiting adaptive coding and modulation,” International Journal of Satellite Communications and Networking, vol. 26, no. 2, pp.

157–188, Mar. 2008. [Online]. Available: https://onlinelibrary.wiley.

com/doi/10.1002/sat.901

[12] “Adaptive Coding and Modulation (ACM) | Comtech EF Data.”

[Online]. Available: https://www.comtechefdata.com/technologies/

acm

[13] R. Acharya, “Chapter 7 - tropospheric impairments: Measurements and mitigation,” in Satellite Signal Propagation, Impairments and Mitigation, R. Acharya, Ed. Academic Press, 2017, pp. 195–245.

[Online]. Available: https://www.sciencedirect.com/science/article/

pii/B9780128097328000077

[14] Acharya, Rajat, “Maxwell’s equation and EM waves,” in Satellite Signal Propagation, Impairments and Mitigation. Elsevier, 2017, pp.

27–56. [Online]. Available: https://linkinghub.elsevier.com/retrieve/

pii/B9780128097328000028

[15] L. Csurgai-Horváth, “Receiver station in Budapest for Q/V band satellite site diversity and adaptive coding and modulation experiments with Alphasat,” International Journal of Satellite Communications and Networking, vol. 37, no. 3, pp. 149–162, May 2019. [Online]. Available: https://onlinelibrary.wiley.com/doi/10.

1002/sat.1270

[16] “ETSI EN 302 307-1 V1.4.1 Digital Video Broadcasting (DVB); Second generation framing structure, channel coding and modulation systems for Broadcasting, Interactive Services, News Gathering and other broadband satellite applications;,”

Nov. 2014. [Online]. Available: https://www.etsi.org/deliver/etsi_en/

302300_302399/30230701/01.04.01_60/en_30230701v010401p.pdf [17] K. P. Murphy, Machine learning: a probabilistic perspective, ser.

Adaptive computation and machine learning series. Cambridge, MA:

MIT Press, 2012.

[18] “Difference between Artificial intelligence and Machine learning - Javatpoint.” [Online]. Available: https://www.javatpoint.com/

difference-between-artificial-intelligence-and-machine-learning [19] T. Rossi, M. De Sanctis, M. Ruggieri, C. Riva, L. Luini, G. Codispoti,

E. Russo, and G. Parca, “Satellite communication and propagation experiments through the alphasat Q/V band Aldo Paraboni technology demonstration payload,”IEEE Aerospace and Electronic Systems Magazine, vol. 31, no. 3, pp. 18–27, Mar. 2016. [Online].

Available: http://ieeexplore.ieee.org/document/7478406/

[20] A. Saffari, C. Leistner, J. Santner, M. Godec, and H. Bischof, “On- line random forests,” in2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops, 2009, pp. 1393–1400.

[21] J. Quinlan, “Simplifying decision trees,” International Journal of Man-Machine Studies, vol. 27, no. 3, pp. 221–234, Sep.

1987. [Online]. Available: https://linkinghub.elsevier.com/retrieve/

pii/S0020737387800536

[22] T. Hastie, R. Tibshirani, and J. H. Friedman,The elements of statistical learning: data mining, inference, and prediction: with 200 full-color illustrations, ser. Springer series in statistics. New York: Springer, 2001.

[23] M. T. Hagan, H. B. Demuth, M. H. Beale, and O. De Jésus,Neural network design, 2nd ed. s.L: Martin T. Hagan, 2014.

[24] I. Goodfellow, Y. Bengio, and A. Courville,Deep Learning. MIT Press, 2016, http://www.deeplearningbook.org.

[25] B. Roy, R. Acharya, and M. Sivaraman, “Attenuation prediction for fade mitigation using neural network with in situ learning algorithm,” Advances in Space Research, vol. 49, no. 2, pp. 336–

350, Jan. 2012. [Online]. Available: https://linkinghub.elsevier.com/

retrieve/pii/S0273117711007435

[26] Á. L. Makara, T. Deli, and L. Csurgai-Horváth, “AI-Supported Fading Prediction,” in 26th Ka Band Communications Conference 2021 Propagation, ser. Propagation 4:. Washington DC: Ka and Broadband Communications, Navigation and Earth Observation Conference, Sep. 2021, pp. 1–5, iSSN-2573-6124. [Online]. Available:

https://proceedings.kaconf.com/papers/2021/ka4_2.pdf

[27] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,”

Neural Computation, vol. 9, no. 8, pp. 1735–1780, Nov. 1997. [Online].

Available: https://direct.mit.edu/neco/article/9/8/1735-1780/6109 [28] M. K. Simon and M.-S. Alouini, Digital communication over fading

channels: a unified approach to performance analysis, ser. Wiley series in telecommunications and signal processing. New York: John Wiley

& Sons, 2000.

[29] K. E. Koech, “Cross-Entropy Loss Function,” Jul.

2022. [Online]. Available: https://towardsdatascience.com/

cross-entropy-loss-function-f38c4ec8643e

[30] D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,”

2014. [Online]. Available: https://arxiv.org/abs/1412.6980