Machine learning aided switching scheme for hybrid FSO/RF transmission

Abstract

Free Space Optics (FSO) is one of the technologies which supports immense data transfer requirements. Though it offers high data rate, but experiences atmospheric attenuation due to dynamic weather conditions. On the other hand, RF communication has lower data rates but are comparatively insensitive to weather conditions. This paper focuses on a hybrid FSO/RF system with the application of Machine Learning (ML) on the prediction of Link Margin (LM) and a ML based switching mechanism between FSO/RF based on the current weather conditions. LM is considered as an important quality parameters in the design and analysis of the FSO link. Mainly rain and fog meteorological data are considered for the estimation and classification of link.

Keywords

FSO link margin atmospheric attenuation machine learning (ML)

1. Introduction

Figure 1.

System model of ML based switching.

FSO is a Line-of-Sight (LoS) technology with a high data rate of nearly 10 Gbps is achieved in a full-duplex communication to transmit various streams of voice, data, high-quality video services etc [1]. As these system serves over free (unlicensed) optical frequencies and tends to work longer distances up to a few kilometers [2, 3] they can avail the everlasting demand for higher data rates. FSO technology uses an optical beam to transmit data in free space, providing optical connectivity. FSO works similarly in operation as optical fiber communication with a difference of channel it propagates, i.e., optical beams are transmitted through atmosphere instead of optical fiber. FSO system employs optical transceiver with bi-directional communication capability at both ends. A few of the essential characteristics of an effective FSO system are [1]: a) To establish a long distance link, the FSO system needs to operate at higher power. b) Transmission of a higher data rate requires a high-speed modulation technique is an important issue to consider. c) FSO systems are to be designed such that low power is consumed, lowering the maintenance of the system. d) Also, it should operate at varied temperature regions (i.e., the outdoor environment should not degrade the performance of the system). e) Mean time between failures (MTBF) must be high enough, i.e., more than ten years is a considerable choice. Moreover, FSO and RF channel illustrate reciprocative properties towards atmospheric and weather effects. In particular, the FSO link is less sensitive to rain [4, 5], whereas the link performance of FSO degrades vitally in case of fog. On the other hand, RF is very much sensitive to rain but not so for fog. Hence a combination of FSO and RF link provides the complementary advantage of both the technologies i.e., high reliability, non-LoS, low latency and high data rate. There are various previous works on hybrid FSO/RF system which focus on soft-switching between two links using the joint design of coding schemes [6, 7, 8, 9]. A hard switching based on the current status of the link, i.e., received power above the threshold is performed [10]. In [7], a rateless coded automatic repeat request (ARQ) scheme has been proposed for hybrid FSO/RF systems. A bit interleaved coded modulation scheme is proposed in [8] for such hybrid systems. In [9] proposed the raptor coding to mitigate the large burst of data due to channel conditions. From an information theory perspective, the link availability of hybrid FSO/RF was examined in [11]. Effect of non-Gaussian noise with similar hybrid FSO/RF system and its performance is analyzed in [12]. A switching scheme based on a single threshold for RF link and the double threshold for the FSO link is proposed in [13]. In [14] switching is based on the signal to noise ratio at the receiver of the FSO system. In [15] designed the switching mechanism based on the single threshold and double threshold with FSO as the highest priority; performance of the system is analyzed. In a recent [16], a predictive switching mechanism based on the deep learning algorithm to optimize the energy is proposed. Some rate loss is observed in hard-switching schemes as compared to soft-switching schemes. To the best of our knowledge, switching mechanisms based on the ML considering the current weather data has not been addressed. This work uses the applications of ML for selecting the communication link based on current atmospheric conditions.

The remainder of the paper is organized a follows. Following to introduction, Section 2 presents the proposed system model of ML-based switching system in hybrid FSO/RF. Section 3 discusses the FSO link design considerations and different atmospheric effects. Section 4 discusses the various ML models in determining link availability and switching accordingly. Section 5 presents the results of the proposed scheme, and finally, Section 6 provides the concluding remarks of the work.

2. System model

The system model of the proposed ML-based switching is shown in Fig. 1. It consists of a hybrid FSO/RF transceiver connected to a switching system. A detailed implementation of the experimental implementation of hard-switching can be obtained in our previous work [10]. The current atmospheric condition from the meteorological department is obtained through Internet of Things (IoT) cloud. ML algorithms are used for LM estimation, and classification as an FSO link endures several effects due to atmospheric attenuation (i.e., losses created due to fog, rain, snow), turbulence, scattering, misalignment, etc. LM estimation provides the switching system to select for data transmission between the FSO or RF link. Along with the LM estimation, classification of data transmission links is determined using the ML model, which thereby controls the switching system. Real-time weather data is provided as input to the LM estimation block.

3. Hybrid FSO/RF link design

3.1 FSO link design

FSO is a technology that transmits the optical beam into the atmospheric channel as a propagating medium. The properties of the channel are the random functions of space and time. Designing of the FSO link is dependent on atmospheric conditions and geographical deployment location. Several unpredictable environmental factors, i.e., fog, rain, clouds, snow, haze, etc., create attenuation on the transmitted optical signal. This limits the link operational distance of the FSO system. Different challenges observed by the system designer in terrestrial FSO links are discussed. Power loss due to atmospheric attenuation on the laser beam is dependent on different parameters. An essential quality parameter in the design is the link margin (LM). It is dependent on different attenuation factors that affect the link. Few of the vital attenuation are explained below:

3.2 Geometrical attenuation ( $\alpha_{\textit{geo}}$ )

Optical receivers have a small capture area where a very less amount of power is received, as the light beam travels for a long distance in free space. This leads to geometrical attenuation given by [17],

$\displaystyle\alpha_{\textit{geo}}=\frac{S_{t}}{S_{c}}$ (1)

where, $S_{t}=\frac{\pi}{4}(d\theta)^{2}$ is the area of illumination. $\theta$ corresponds to divergence angle of the beam; $d$ is the distance from transmitter to receiver. $S_{c}$ is the capture area of the receiver.

3.3 Atmospheric attenuation (

\alpha_{\textit{atmo}}

)

Due to variation in weather conditions, there are different types of atmospheric attenuation (i.e., rain, fog, snow, turbulence, etc.).

3.3.1 Rain attenuation

The impact of rain is dependent on the intensity of rain. The attenuation created for light rain (2.5 mm/hr) to heavy rain (25 mm/hr) ranges from 1 dB/km to 10 dB/km for wavelengths around 850 nm and 1500 nm [18]. Attenuation due to rain can be represented as,

$\displaystyle\alpha_{\textit{rain}}={1.076}\ast R^{0.67}$ (2)

where, $R$ is the rain intensity in mm/hr.

Table 1

Different values of $q$

$q$	$v$
1.6	if $v>$ 50 km
1.3	if 6 km $<v<$ 50 km
0.16	if 1 km $<v<$ 6 km
$v$ -0.5	if 0.5 km $<$ 1 km
0	0.5 km $<v$

3.3.2 Fog attenuation

Attenuation created due to fog has a crucial role on the FSO link. Attenuation is observed due to absorption and scattering from fog particles. Typical dense fog condition limits the visibility to less than 50 m creating an attenuation of more than 350 dB/km [19]. This limits the availability of the FSO link. Fog attenuation in terms of visibility can be expressed as,

$\displaystyle\alpha_{\textit{fog}}(\lambda)=\frac{3.91}{v}\left(\frac{\lambda}% {550\ast 10^{-9}}\right)^{-q}$ (3)

where, $q$ values with respect to visibility ( $v$ ) is mentioned in Table 1. Apart from the rain and fog, snow attenuation, scintillation effect, misalignment losses, etc., are also essential. However, this work focuses on the attenuation created due to rain, fog, and not limited, but a similar analysis can be extended to other events. An FSO link availability is dependent on LM, based on different attenuation of Eqs (1)–(3) LM can be expressed as follows,

$\displaystyle LM=P_{e}+|S_{r}|-\alpha_{\textit{atmo}}-\alpha_{\textit{geo}}-% \alpha_{\textit{sys}}$ (4)

where, $P_{e}$ is the transmitter power; $|S_{r}|$ is the receiver sensitivity; $\alpha_{\textit{geo}}$ is geometrical attenuation; $\alpha_{\textit{sys}}$ is system losses and $\alpha_{\textit{atmo}}$ is atmospheric attenuation which is a collective representation of $\alpha_{\textit{rain}}$ and $\alpha_{\textit{fog}}$ . Detailed performance of the FSO link under different atmospheric condition is evaluated in our previous work [17].

3.4 RF link design

The complimentary nature of FSO and RF has proven to emanate as a practical solution for reliable connectivity with a high data rate. In [4] discusses the detail effects of weather in the RF communication link.

4. Link quality estimation and prediction using machine learning algorithms

In a hybrid, FSO/RF system, estimating the quality of the link is based on the free space conditions is a challenging task. Also, switching between the two communication links based on the current atmospheric requires a real-time monitoring and control system. In order to accomplish this, the following are the different ML algorithms considered for evaluation.

4.1 Logistic regression

Logistic regression is used when a dependent variable is based on one or more independent predictive variable, which predicts the outcome. Input arguments (i.e., $x$ ) are linearly combined with weights to predict the outcome ( $y$ ). The main difference with linear regression is the output, which is modeled as binary (i.e., 0 or 1) rather than a numeric value. A typical logistic regression can be represented as [20],

$\displaystyle y=\frac{e^{(b_{0}+b_{1}\ast x)}}{(1+e^{(b_{0}+b_{1}\ast x)})}$ (5)

where, $b_{0}$ represents the bias, $b_{1}$ the associated weight of the single input ( $x$ ) and $y$ is the output to be predicted. Input data vector has to be associated with weight vector which is to be learned from the training data.

4.2 K-nearest neighbor

$K$ -nearest neighbors is a non-linear algorithm that classifies the cases based on the closest similarity measures (e.g., distance functions). More the neighbors, the more chance to find correct predictions, and therefore accuracy has the possibility to increase. However, too many neighbors can cause overfitting on the training set, and the predictions will be poor on new observations in the test set. One of the prominent distance function used in the above algorithm is Euclidean distance which can be expressed as [21],

$\displaystyle D=\sqrt{\sum_{i=1}^{k}}(x_{i}-y_{i})^{2}$ (6)

where, $x_{i}$ is the query instance and $y_{i}$ is the training sample.

4.3 Naive Bayes

Naive Bayes algorithm classification is based on Bayes theorem with naives assumptions. Its learning assumes that features are independent, but classifies very well. It is a sub-set of bayesian decision theorem, which can be expressed as [21],

$\displaystyle P\left(\frac{A}{B}\right)=\frac{P\left(\frac{B}{A}\right)P(A)}{P% (B)}$ (7)

where, $P(A)$ represents the probability of event $A$ , $P(B)$ , probability for the event $B$ to occur, $P(A|B)$ is the conditional probability where $A$ , given that $B$ occurs. $P(B|A)$ the conditional probability where $B$ given that $A$ occurs. It is also referred to as likelihood.

4.4 Random forest

The random forest can be the well suitable algorithm for classification problems. It is a supervised learning algorithm that can be used for both classification and regression by considering the multitude of decision trees with random training samples. Prediction is performed based on the voting to select the best solution. It provides better predictive power than the decision trees. Averaging the results of different decision trees, random forest overcomes the overfitting problem and provides good accuracy [22].

5. Results and discussion

Based on the system model and mentioned ML algorithms, LM of the FSO system is predicted for switching between FSO or RF link. The performance of a machine learning model is determined with metrics such as confusion matrix, precision, recall, and entropy. Confusion matrix values are obtained to analyze the performance of different ML algorithms. Accuracy or classification rate is an important metric in determining the performance of the ML algorithms. Accuracy is calculated from the obtained confusion matrix values given by [23],

$\displaystyle\textit{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN}$ (8)

where, $T P$ : True Positive, $T N$ : True Negative, $F P$ : False Positive, $F N$ : False Negative. Two of the atmospheric conditions i.e., rain and fog conditions have been considered. A detailed implementation is as follows:

Table 2

Data set used for rain simulation

Time stamp	Rain fall in mm		Attenuation (dB/km)
2:24:00	1	.5	1.41186
3:50:00	2		1.71149
5:50:00	4		2.72390
16:30:00	10		5.03283
22:20:00	14	.5	6.45548
1:15:00	22		8.53562

5.1 Rain attenuation

Attenuation produced due to rain has been predicted using ML algorithms. The machine is trained using a real-time data set obtained from the Earth and Atmospheric Sciences department of NIT Rourkela, which calibrates the rainfall rate at an interval of one minute. A few of the observations from the data set with rainfall at different time stamps are shown in Table 2. Linear regression is initially used for predicting the rain attenuation. The model is trained with 75% of the data and tested for the remaining 25%. Figure 2 shows the prediction of rain attenuation for different precipitation of rain using linear regression. For arbitrary rain precipitation of 7.5 mm, the predicted attenuation is 3.6823 dB/km by the model, and on calculating using Eq. (2), it is 4.105 dB/km, with an error of 0.4227. To get accurate predictions of LM, different ML models discussed in Section 4 are applied. Figure 3 represents the training model of logistic regression with decision boundaries based on LM and rainfall rate. Figure 4 represents the testing phase to classify the LM, red region, i.e., ‘0’ indicates the RF link to be used and the green region as the FSO link. Figures 5 and 6 represents the random forest classifier model implementation. Similarly, other ML algorithms, i.e., K-NN, Naive Bayes classifier, are applied.

Figure 2.

Prediction of rain attenuation using linear regression.

Figure 3.

Training of logistic regression for prediction of different rain attenuation.

Figure 4.

Prediction of LM for different rain intensity using logistic regression.

Figure 5.

Training of random forest classifier different rain intensity.

Figure 6.

Classification of LM for different rain intensity using random forest classifier.

Table 3

Machine learning algorithms and their corresponding accuracy for rain attenuation

Machine learning model	Accuracy (in %)
Logistic regression	90.0
K-NN classifier	98.0
Naive Bayes	96.8
Random forest	100

The accuracy of prediction for ML algorithms applied is shown in Table 3. Random forest algorithm outperforms than other algorithms in predicting LM of the FSO system under rain atmospheric.

5.2 Fog attenuation

Table 4
Data set used for link selection

Visibility (km)	$q$	Link Margin (dB)	FSO/RF
0.1	0	$-$ 27.03	0
1	0.5	12.06993	1
4	0.16	12.03993	0
25	1.3	12.07	1
60	1.6	12.07	1
100	1.6	12.07	1

Figure 7.

Training of logistic regression to predict LM with visibility values of fog.

Figure 8.

Prediction of LM due to fog attenuation using logistic regression.

Figure 9.

Training of K-NN classifier to predict LM with visibility values of fog.

Figure 10.

Prediction of LM due to fog attenuation using K-NN classifier.

Figure 11.

Training of Naive Bayes classifier to predict LM with visibility values of fog.

Figure 12.

Prediction of LM due to fog attenuation using Naive Bayes classifier.

Figure 13.

Training of random forest classifier to predict LM with visibility values of fog.

Figure 14.

Prediction of LM due to fog attenuation using random forest classifier.

The measurement of fog attenuation is based on the visibility data. Fog visibility data is used to train the model to predict the fog attenuation and corresponding LM based on Eqs (3) and (4). ML model needs to classify the LM with threshold and provide the information regarding the active link of hybrid FSO/RF to the switching system. A predefined threshold of 12.05 dB is used to classify the link. A few of the readings of the data set used for the simulation are mentioned in Table 4. After training, the ML model predicts the link margin for given visibility. Figure 7 represent the training of logistic regression with the red region as RF link and the green region as the FSO link. Figure 8 shows the testing phase with the logistic regression model where the accuracy of prediction is 65% and 35% are misclassified, which has a poor performance. Figure 9 is the training of the K-NN classifier, which is a non-linear model with high prediction based on the neighbors. Figure 10 represents the classifier testing phase, which has good prediction and classification. Naive Bayes classifier is trained to classify the LM is shown in Fig. 11. Figure 12 shows the testing of the model, Naive Bayes classifier has good performance compared to the logistic regression, but K-NN has a better prediction compared to Naive Bayes. Also, a random forest ML model is applied with the training phase is shown in Fig. 13. Figure 14 represents the testing phase, where it has a superior performance compared to other ML models. The performance of the ML models in terms of accuracy, is shown in Table 5. It is observed that the random forest algorithm outperforms the other ML algorithms in predicting and classification of link.

Table 5

Machine learning algorithms and their corresponding accuracy

Machine learning model	Accuracy (in %)
Logistic regression	65.0
K-Nearest Neighbour classifier	95.0
Kernel SVM	70.0
Naive Bayes	90.0
Random forest	100

6. Conclusion

This work highlights the switching mechanism based on the machine learning algorithm in a hybrid FSO/RF system. Real-time weather data is taken into consideration in determining the link quality. ML algorithm predicts the LM of an FSO link and classifies the communication link to have opted under current channel conditions. Models are trained with real-time rain, and fog data set to determine the link quality. Among the other ML models, random forest has shown superior performance in terms of accuracy in predicting and classification. Experimental implementation of the ML model can provide more insight into the proposed scheme, and it forms the part of future research work.

Footnotes

Acknowledgments

This work has been supported by Department of Electronics and IT (DeitY), Ministry of Communications and IT, Government of India under Visvesvaraya PhD Scheme for Electronics and IT (Grant no: PhD-MLA/4(13)/2015-16).

References

Chan

. Free-space optical communications. Journal of Lightwave Technology. 2006; 24(12): 4750–4762.

Andrews

Phillips

Hopen

. Laser beam scintillation with applications. vol. 99. SPIE press; 2001.

Giannetti

Luise

Reggiannini

. Mobile and personal communications in the 60 GHz band: a survey. Wireless Personal Communications. 1999; 10(2): 207–243.

Nadeem

Kvicera

Awan

Leitgeb

Muhammad

Kandus

. Weather effects on hybrid FSO/RF communication link. IEEE Journal on Selected Areas in Communications. 2009; 27(9): 1687–1697.

Kavehrad

. Availability evaluation of ground-to-air hybrid FSO/RF links. International Journal of Wireless Information Networks. 2007; 14(1): 33–45.

Vangala

Pishro-Nik

. A highly reliable FSO/RF communication system using efficient codes. In: IEEE GLOBECOM 2007-IEEE Global Telecommunications Conference. IEEE; 2007. pp. 2232–2236.

AbdulHussein

Oka

Nguyen

Lampe

. Rateless coding for hybrid free-space optical and radio-frequency communication. IEEE Transactions on Wireless Communications. 2010; 9(3): 907–913.

Schober

. Bit-interleaved coded modulation for hybrid RF/FSO systems. IEEE Transactions on Communications. 2009; 57(12): 3753–3763.

Zhang

Hranilovic

Shi

. Soft-switching hybrid FSO/RF links using short-length raptor codes: design and implementation. IEEE Journal on Selected Areas in Communications. 2009; 27(9): 1698–1708.

10.

Kiran

Rathore

Turuk

Das

. Development of a Hybrid FSO/RF System during Link Misalignment. In: 2017 International Conference on Networking and Network Applications (NaNA). IEEE; 2017. pp. 138–140.

11.

Letzepis

Nguyen

Fàbregas

Cowley

. Outage analysis of the hybrid free-space optical and radio-frequency channel. IEEE Journal on Selected Areas in Communications. 2009; 27(9): 1709–1719.

12.

Kamboj

Mallik

Agrawal

Schober

. Diversity combining in FSO systems in presence of non-Gaussian noise. In: 2012 International Conference on Signal Processing and Communications (SPCOM). IEEE; 2012. pp. 1–5.

13.

Shrivastava

Sengar

Singh

. A new switching scheme for hybrid FSO/RF communication in the presence of strong atmospheric turbulence. Photonic Network Communications. 2019; 37(1): 53–62.

14.

Sharma

Madhukumar

Swaminathan

. Switching-based hybrid FSO/RF transmission for DF relaying system. In: 2018 IEEE Wireless Communications and Networking Conference (WCNC). IEEE; 2018. pp. 1–6.

15.

Usman

Yang

Alouini

. Practical switching-based hybrid FSO/RF transmission and its performance analysis. IEEE Photonics Journal. 2014; 6(5): 1–13.

16.

Meng

Liu

Song

Yang

Guo

. Predictive Link Switching for Energy Efficient FSO/RF Communication System. In: Asia Communications and Photonics Conference. Optical Society of America. 2019. pp. T4C–1.

17.

Kiran

Kumar

Turuk

Das

. Quality Aware Design and Analysis of FSO Link in Wireless Optical System. In: 2017 14th IEEE India Council International Conference (INDICON). IEEE; 2017. pp. 1–6.

18.

Kim

Achour

. Free-space links address the last-mile problem. Laser Focus World. 2001; 37(6): 121–130.

19.

Willebrand

Ghuman

. Free space optics: enabling optical connectivity in today’s networks. SAMS publishing; 2002.

20.

Hilbe

. Logistic regression models. Chapman and hall/CRC; 2009.

21.

Obulesu

Mahendra

ThrilokReddy

. Machine Learning Techniques and Tools: A Survey. In: 2018 International Conference on Inventive Research in Computing Applications (ICIRCA). IEEE; 2018. pp. 605–611.

22.

Belgiu

Drăguţ

. Random forest in remote sensing: a review of applications and future directions. ISPRS Journal of Photogrammetry and Remote Sensing. 2016; 114: 24–31.

23.

Hossin

Sulaiman

. A review on evaluation metrics for data classification evaluations. International Journal of Data Mining & Knowledge Management Process. 2015; 5(2): 1.