Open Access
Add to BibSonomy
Abstract
Free space optical (FSO) communication has attracted significant attention due to its high transmission rate and information security. Nevertheless, the FSO link is sensitive to various weather conditions which limit its application range. Thus, research on the FSO channel plays an important role for combatting channel fading. In the paper, we first establish a FSO transmission testbed to obtain received signal strength indication (RSSI) information of optical signal. Then we utilize an environmental chamber for indoor experiments to simulate weather changes in the real world. Finally, based on the true meteorological dataset from the official department and the RSSI dataset from the experiments, we employ expanded gated recurrent unit (GRU) neural networks to implement the FSO channel prediction. The results demonstrate that the proposed scheme can achieve the prediction of FSO channel fading with a high precision, where the absolute percentage error (APE) values lower than 6.9% account for up to ninety% of results.
© 2021 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
1.1 Background
With the explosion of demands for smart devices and network traffic over the past few years, a higher data rate technique is promised and thus free space optical (FSO) communication comes into academic and industrial views [1–3]. FSO communication can provide high-capacity transmission, which is up to 1.2 Tbps utilizing new optical modules [4]. Compared with conventional Radio Frequency (RF) technology, FSO communication has advantages of inherent high security, flexibility, rapid deployment and electromagnetic immunity. Although optical fiber communication is still the main technology of high-speed transmission, FSO communication technology will bring its superiority into full play in the areas that fiber is inaccessible. Besides, FSO communication also has other applications such as front-haul of the cloud radio access network, emergency communication and tactic communication [5,6].
FSO links need line-of-sight transmission without obstructions in the range of sight. Nevertheless, FSO link is susceptible to atmospheric turbulence and weather conditions, which can reduce channel transmission quality and limit related applications in long-distance communication. Thus, more and more attention has been paid to the channel research of FSO communication in recent years. At present, there are many researches to build various turbulence models coping with turbulence conditions of different intensity [7–9] and also many channel attenuation statistical models for different weather channels [10–12]. Related studies about turbulence and various weather channels have important guidance for the FSO channel transmission models. According to the proposed models, the outage probability and system performance have been further optimized and analyzed.
Most related works are devoted to channel evaluation under current channel conditions but do not pay attention to the changing trend at the future moment. However, for transmission stability and adaptability, it is of great significance to obtain the changing trend of channel conditions in the future. In order to achieve FSO channel prediction effectively, this paper proposes a Machine Learning (ML) assisted prediction scheme to support the intelligent FSO transmission system. ML techniques have been widely studied in optical communication for fiber nonlinear effect mitigation [13], performance monitoring [14] and modulation format identification [15,16]. However, most researches are focused on optical-fiber communication and few consider FSO applications. It is worth mentioning that relevant researches are still in the earlier stage and there are still a lot of research opportunities. For example, [17] employed the Support Vector Machine (SVM) as a detector to combat the noise and scintillation effects. [18] used Conventional Neural Network (CNN) as an adaptive demodulator for Turbo-coded 16-ary Orbital Angular Momentum (OAM) shift keying FSO communication system. In [19], CNN was used for joint atmospheric turbulence detection and demodulation for an OAM FSO system. A novel deep-learning based detector was proposed for FSO system employing a Deep Neural Network (DNN) for detection without CSI [20]. The Back Propagation (BP) artificial neural network was employed for the sensor-less AO system to design a distortion correction scheme [21]. There are some other researches focused on system modeling [22,23], and geometric shaping [24]. Correspondingly, ML techniques can be employed in FSO channel estimation and models.
1.2 Contributions
Based on plentiful experimental studies, we found and verified the predictability of FSO communication channels. Although the fact that FSO signal is vulnerable to weather conditions including fog, snowing, heavy rain, haze, etc [12], Received Signal Strength Indication (RSSI) of optical signal reflects the transmission quality directly. Thus, we use RSSI to characterize the channel variation. In the paper, we accomplish the goal of analyzing and estimating channel state changes by predicting RSSI with neural networks. The contributions of this paper are fourfold:
- • We establish a FSO transmission testbed with intensity modulation (IM)/direct detection (DD) and the transmission rate is up to 10Gbps. Based on the testbed, we further develop the data acquisition system for RSSI monitoring and collection.
- • Through National Meteorological Information Center (NMIC), we obtain the environmental parameters in real world. Based on part of parameters, we make a database for ten years and then carry out plenty of repeatable experiments to simulate channel conditions by building an environment chamber.
- • Based on massive channel transmission experiments, we get a RSSI dataset. By analyzing the RSSI dataset, we find a certain autocorrelation with regards to RSSI, which implies that the RSSI is predictable. We further formulate the FSO channel prediction as a RSSI time series prediction problem and propose a rolling prediction scheme by combining gated recurrent units (GRU) neural network.
- • On basis of abundant experimental results and analysis, the proposed scheme can effectively predict the changing trend of FSO channel. For example, in the experiment of 4m transmission distance, the minimum mean absolute error (MAE) can reach about 0.3dBm and absolute percentage error (APE) values less than 6.9% can account for 90% of all prediction results.
The structure of this paper is organized as follows. The testbed setup and environmental chamber are introduced in Section 2. Section 3 presents the proposed rolling channel prediction scheme using ML algorithm. Then in Section 4, the collection, processing and analysis of environmental dataset and RSSI dataset are described. Section 5 shows experimental results and discussions. Finally, conclusion is given in Section 6.
2. Experimental setup
The setup of FSO transmission system is shown in Fig. 1. FSO links are composed of two FSO transceivers, which are placed on either side of the meteorological environment simulating chamber (MESC).
2.1 FSO system setup
Communication link: The established testbed works in full-duplex. However, for the purpose of clarity, we show a unidirectional transmission link in the Fig. 1. The electrical signal is generated by arbitrary waveform generator (AWG), and then a Mach-Zehnder modulator (MZM) is used to complete the intensity modulation (IM) of the distributed feedback laser (DFB) with 1550nm wavelength. Next, the modulated optical signal goes through erbium-doped fiber amplifier (EDFA) module, which amplifies the optical signal and emits it through the lens. In the system, it is worth mentioning that there are 2×1 links used for data transmission to obtain the gain of spatial diversity and combat turbulence. Two laser links transmit the same optical signal and FSO receivers can get them by the same lens. At the receiving side, the two beams of laser can be received by the lens and then split the signal strength in a ratio of 1 to 19. The laser beam with 95% optical power is amplified by EDFA and then used to transmit information, while another laser beam with 5% optical power is detected by photodiode detector (PD) to measure RSSI. Finally, signal is transmitted to oscilloscope and the demodulated signal waveform can also be displayed. The transmission rate of the FSO system is up to 10Gbps.
FSO control and optical power detection system: FSO control system is composed of Upper Computer Control System (UCCS) and Lower Computer Control System (LCCS). UCCS can transmit data and control the LCCS via serial-port communication. After passing through the IM module, optical signal is processed by EDFA module and then emitted through lens. At the receiving side, the optical power detector will test the RSSI of optical signal and then RSSI is fed back to UCCS. According to the obtained RSSI dataset, the proposed channel prediction scheme is achieved and integrated into UCCS.
2.2 MESC system setup
Figure 2(a) presents the structure of MESC. The main body of MESC is a hexahedral closed chamber, which measures 400 centimeters long, 240 centimeters wide and 330 centimeters high. There are two glass windows coated with high transmittance film for the laser to pass through, the size of which are 30 centimeters long and 30 centimeters wide. Users can input environmental parameters to implement comprehensive control of the electromechanical equipment and further achieve the environmental conditions that users desire. MESC can be used for simulations of natural illumination, temperature, humidity, partial pressure, wind speed, rainfall, fog, haze and snowfall. In the paper, fog condition is selected as the attenuation of laser transmission, which is generated by ultrasonic humidifier and evaluated by visibility tester. Figure 2(b) shows the photographs of FSO transmission experiments in fog condition. The related technical parameters are as follows:
- • Temperature: control range −20–60°C, control accuracy ≤ ±0.2°C, uniformity ≤ ± 0.5°C;
- • Humidity: control range 30–95%RH, control accuracy ≤ ±3%RH, uniformity ≤ ±3%RH;
- • Fog: 0–30000m (continuously adjustable), uniformity ≤ ±3%.
2.3 MESC database setup
To simulate environmental changes efficiently, we also develop UCCS and LCCS for MESC. The UCCS of MESC has access to the database. The environmental parameters set by UCCS are sent to LCCS in an online way and MESC can produce the corresponding operating conditions. We selected five representative cities from five geographical locations with different meteorological characteristics in China for the experimental analysis, including Chongqing (106.5°E, 29.6°N), Shanghai (121.4°E, 31.2°N), Lhasa (91.1°E, 29.6°N), Shenyang (123.4°E, 41.7°N) and Beijing (116.4°E, 40.4°N), as shown in Fig. 3. Chongqing, located in the southeast of Sichuan Basin, has a subtropical monsoon climate with more fog and less sunshine throughout the year. Shanghai has a subtropical maritime monsoon climate with moderate rainfall and a relatively even seasonal distribution. Lhasa is located in the northern side of the Himalayas, with little rainfall and long hours of sunshine throughout the year. Beijing’s climate is featured by a typical north temperate sub-humid continental monsoon, with high temperature and rain in summer. Shenyang also has a temperate semi-humid continental climate with concentrated precipitation, large temperature difference and four distinct seasons. FSO link is susceptible to obscuring obstacles and weather conditions. Since the size of fog is similar to the laser wavelength, the effect of fog on the FSO link is especially obvious. In nature, rain and fog often appear together and thus when the local weather is cloudy with rain and fog, the FSO link will be severely attenuated and the RSSI will be significantly decreased. We obtain the meteorological data of the five cities in China from NMIC. Then, we create the meteorological database of the five cities from 2010 to 2019. The data content includes geographical location, altitude, temperature, humidity, pressure, precipitation, wind speed, wind direction and visibility. The database consists of a total of 6,071,160 items of data and includes a total of 115,352,040 parameter values.
3. Expanded GRU neural network for FSO channel with rolling prediction
Recurrent Neural Network (RNN) is used to memory the long-term dependency [25]. However, with the increasing time lags, gradients of networks may vanish, which significantly reduces network performance. Thus, new structures of RNN such as long short-term Mmemory (LSTM) and GRU have been proposed to make memory cells determine when and how much to forget previous information [26]. In the paper, the GRU neural network is extended to a rolling prediction scheme for FSO channel.
GRU is a variant of LSTM, introduced by K, Cho and considered to be much easier to compute and implement than LSTM [27]. GRU retains the resistance of LSTM to the vanishing gradient problem. However, the structure of GRU is simpler and therefore easier to train for the reason of less computation needed to update hidden states. According to Fig. 4, GRU has update gate and reset gate, which are similar to the function of forget gate and input gate in LSTM. The update gate determines how much of the previous memory to be remembered. The reset gate decides how to combine the new input with the previous memory. The main difference is that the GRU only uses integrals to fully disclose its memory contents.
In GRU neural network, the update gate state ${z_t}$ determines how many states ${x_t}$ are retained and how many candidate hidden states ${h_{t - 1}}$ can be kept in current output state information. ${W_{xz}}$ and ${W_{hz}}$ are the weighting parameters. ${b_z}$ is the deviation parameter. The update gate is calculated as Eq. (1) [28].
Reset gate state decides if the candidate state information at the present moment depends on the past state information ${h_{t - 1}}$ and the dependency level ${r_t}$ can also be determined. ${W_{xr}}$ and ${W_{hr}}$ are the weighting parameters. ${b_r}$ is the deviation parameter. The reset gate is calculated as Eq. (2) [28].
The value of ${r_t}$ determines the degree of dependency of the candidate hidden state ${h^{\prime}_t}$ on the state ${h_{t - 1}}$ at the past moment. Figure 4 shows that the state information at the previous time is first multiplied by the output of the reset gate and then further used as the parameter to calculate the candidate state information at the current time. ${W_{xh}}$ and ${W_{hh}}$ are the weighting parameters. ${b_h}$ is the deviation parameter. The formula is shown in Eq. (3) [28].
The output of the update gate is multiplied with the candidate hidden state ${h^{\prime}_t}$. The historical candidate hidden state ${h_{t - 1}}$. is multiplied by $1 - {z_t}$. The output of the neural network at the current moment is calculated in Eq. (4) [28].
Figure 5 shows the main structure of the proposed prediction model. The input data consist of $RSS{I_{truth}}$, which represents true RSSI values. And the output data consist of $RSS{I_{predicted}}$, which represents predicted RSSI values. In the proposed FSO channel prediction model, a neural network structure of [1, 50, 100, 100, 1] is employed, in which there is 1 input layer (consisting of a sequence of size 50) connected with one GRU layer of 50 neurons. Then the GRU layer is in turn fed into another two GRU layers with 100 neurons further fully connected to a normal layer of one neuron with a linear activation function. Finally, the output layer executes prediction of the next time step.
Algorithm 1.RSSI prediction scheme
In order to make proposed scheme more realistic, we extend the original prediction model to the rolling prediction model. The pseudocode is shown in Algorithm 1. Figure 6 further illustrates the schematic diagram of rolling prediction by taking an example. The window size is set to 50, which is first composed of 50 true RSSI values. Based on the 50 true RSSI values, a predicted RSSI value can be obtained. Then, the predicted RSSI value will be inserted into the prediction window. The data in the prediction window consists of 49 true RSSI values and one prediction RSSI value. As described above, the predicted RSSI values for the second time step can be obtained. The predicted ${R_{n - step}}$ for the ${n_{th}}$ time step will be obtained in turn and formed the predicted sequence.
4. RSSI experiments and data processing
As mentioned above, fog has a huge impact on FSO links and the reason is that the size of fog particle is similar to the wavelength of optical and near infrared wave [29]. Fog can reduce visibility to several meters with attenuation up to 480dB/km in worst cases [30]. Therefore, we mainly consider fog as channel conditions in the paper. We process the meteorological data of January to March in 2019 of 5 cities mentioned above and the time interval of the data is 1 hour. We set visibility and time as matrix $V = {[{T_1},{T_2},\ldots ,{T_n};{V_1},{V_2},\ldots ,{V_n}]^T}$. The parameters are input into UCCS of MESC and the equipment can generate the corresponding channel conditions. The RSSI dataset can be captured at the FSO receiver, which is organized into matrix $R = {[RSS{I_1},RSS{I_2},\ldots ,RSS{I_n}]^T}$. For the sake of less redundant experimentation, we selected two cities with relatively greater fluctuation in visibility in 2019. Figure 7 shows the visibility data in Shenyang and Chongqing as examples, and the resulting RSSI data is shown in Fig. 8(a) and Fig. 8(b), respectively. The curves of two cities display strong volatility and also show slightly different trends in terms of visibility. During January to March in 2019, it can be seen from Fig. 8 that the overall values for visibility of Shenyang are higher and thus the corresponding RSSI is also higher than Chongqing. We can find that the curves of Chongqing’s data in visibility and RSSI are more volatile than Shenyang. It is worth mentioning that in the experimental results of 4m transmission, the lowest value of the RSSI in Shenyang is around −9dBm, while in Chongqing, the lowest value is around −20dBm. The above situations indicate that foggy and humid climate of Chongqing can have a greater impact on the quality of the FSO link.
In the indoor experimental testbed, the link distance between FSO transceivers is 4m. To simulate outdoor channel transmission conditions for a further evaluation, we get more dataset by extending the matrix R to more matrices with different transmission distances according to Kim model [31]. In Section 5, we will correlate both the original and the expanded matrices to ensure the validity of the solution.
In fog condition, Mie scattering is considered as the dominant scattering process in FSO system. The attenuation $\beta $ due to fog is given by Eq. (5) [31]:
Fig. 8. (a). RSSI in Shenyang (b). RSSI in Chongqing (c). RSSI in Shenyang with various transmission distances
For a FSO communication link, the received optical power ${P_r}$(w) is given by Eq. (7) [31]:
Since the relevant parameters of the system are known based on Table 1, ${P_r}$ can become the function of L. Convert the units of the matrix R and the expanded matrices can be obtained. During the expansion of RSSI, a certain amount of error is definitely introduced. However, the purpose of the treatment is to obtain a richer dataset about RSSI. We can better validate the model presented in Section 3 and thus the introduced error is tolerable. To facilitate the display of results, we only take Shenyang as an example. Figure 8(c) shows the RSSI at different transmission distances in Shenyang. As can be seen from the figure, the curves of RSSI fluctuate more sharply as the transmission distance increases. The dataset corresponding to the curves will be used for the prediction experiments in Section 5.
Table 1. FSO transmission parameters
In order to verify the predictability of data, based on the time series analysis of the data, we further consider the Autocorrelation Coefficient (AC) with partial lag order k of the expanded RSSI data. The AC is calculated by Eq. (8) where $RSS{I_t}$ denotes time series of RSSI, N is the time series length and $\mu$ is the mean of $RSS{I_t}$ [32].
Figure 9(a) shows the relationship between the AC and the lag order considering RSSI obtained within the transmission distance of 4m. It can be seen that the correlation of experimental RSSI in different cities is slightly different, but the overall trend is similar. To facilitate the display of results, Shenyang is presented as an example shown in Fig. 9(b). We consider the number of prediction steps in terms of 24 hours in a day as a time period. The lag order is considered as a reference value for prediction steps. Without loss of generality in statistics, we regard a correlation coefficient between 0.1 and 0.3 as a weak correlation. Figure 9(b) shows that the AC values are greater than 0.2 at different transmission distances when the prediction steps are lower than 24. Therefore, we use 24 hours as the maximum prediction steps for the following experimental results to ensure the data correlation and the prediction accuracy.
5. Results and discussion
In this section, we display experimental results for the proposed FSO channel prediction model. The parameters of FSO transceivers are listed in Table 1. We first analyze the relationship between visibility and RSSI based on the datasets obtained in Section 4. Then the corresponding FSO channel prediction results for various scenarios are demonstrated. To facilitate the display of results, the analysis of the proposed scheme in this section still takes Shenyang as an example.
Figure 10 shows the relationship between RSSI and visibility in Shenyang from January 1, to March 31, 2019. The figure ranks the visibility values from lowest to highest. We can observe that the values don’t vary much at the same visibility. Under high visibility conditions up to 10,000 meters, the RSSIs are stable and the values are around −6.5dBm which fluctuates between −0.5dBm and 0.5dBm. When the visibility is below 10,000 meters, it shows that the value decreases gradually, but the trend is not a steep drop in curve due to the limited transmission distance. However, even at such a short transmission distance of 4 meters, it is obvious that the values of RSSI decrease by about 3dBm. When the visibility is around a few hundred meters, the values will drop significantly.
Figure 11 presents MAE of RSSI with different prediction steps. Since the time interval of RSSI is 1 hour, 1 prediction step corresponds to 1 hour. It can be seen from the figure that errors are generally on an increasing trend with the increasing number of predicted steps. The reason is that the weight of the predicted values in the dataset used to make predictions increase constantly, which leads to inaccurate prediction results. As the transmission distance increases, the error values also show an increasing trend, which is due to the greater difficulty in prediction brought by the more dramatic fluctuation of RSSI with longer transmission distance. For example, when the prediction step is 1, the MAE values at 0.004km, 0.2km and 0.5km are 0.2684dBm, 0.2912dBm and 0.3399dBm, respectively. Correspondingly, when the prediction steps are 24, the MAE values at 0.004km, 0.2km and 0.5km are increased by 0.4896dBm, 0.5609dBm and 0.6506dBm, respectively. From the above results, we can calculate that the MAE growth rates at 0.004km, 0.2km and 0.5km are 82.41%, 92.62% and 91.41% respectively. However, the prediction error increases within an acceptable range for a limited number of prediction steps on the whole. We also find that the average error curve fluctuates between adjacent prediction steps and is not a simple linear growth, which is due to the large randomness of the data itself and the different effects of each training session. Therefore, the prediction performance can’t be judged simply by the growth of the number of prediction steps.
Considering different transmission distance, Fig. 12 further presents the error percentage of all prediction results with 10 predication steps from March 27 to March 31, 2019, which means that there are ten prediction periods and a total of 100 prediction data. It can be seen from the figure that the curves fluctuate between various prediction periods and most of the errors are concentrated at less than five percent. The proposed prediction model does not show a significant growth in performance for error percentage with the increasing transmission distance. On the contrary, in terms of error percentages, the six transmission distances show similar performance, which indicates that the proposed model is still valid for the RSSI prediction at long transmission distance with the expanded RSSI dataset based on Eq. (8).
In order to analyze the probabilistic distribution of the results in Fig. 12, we plot the Cumulative Distribution Function (CDF) about Absolute Error (AE) and APE in Fig. 13(a) and Fig. 13(b), respectively. The AE is calculated by $|{RSS{I_{predicted}} - RSS{I_{truth}}} |$ and the APE is calculated by ${{|{RSS{I_{predicted}} - RSS{I_{truth}}} |} / {RSS{I_{truth}}}}$. The CDF presents the percentage of AE and APE data lower than any given value. For example, in Fig. 13(a), we can observe that the AE data lower than 0.4253, 0.5057 and 0.5748 account for seventy percent, respectively. In Fig. 13(b), the AE data lower than 0.8621, 1.02 and 1.538 account for seventy percent, respectively. In Fig. 13(b), the APE data lower than 6.926%, 10.49% and 13.19% account for ninety percent, respectively. The results show that the prediction error of the proposed algorithm grows with the increase of transmission distance, however, it still keeps considerable prediction accuracy.
Figure 14 shows the true and predicted values for RSSI from March 27 to March 31, 2019, where (a), (b) and (c) correspond to the transmission distances of 4m, 100m and 400m, respectively. The prediction step is 10 steps and there are also ten prediction periods. The blue line and the black line represent the true value and the predicted value, respectively. Above or below each prediction period, we mark the corresponding number of periods (i.e., i) and the MAE of the prediction period in red font. Results show the proposed prediction model can efficiently and accurately predict the RSSI trend. The prediction algorithm does not show large fluctuations in prediction performance with changes in the RSSI dataset. In Fig. 14(a), the fourth prediction cycle has a large deviation. And in Fig. 14(b), the third prediction cycle has a large deviation. Although the values for predicted RSSI have deviation in the individual interval, the overall effect is satisfactory. The prediction accuracy decreases appropriately as the transmission distance increases, while the overall prediction performance remains good.
6. Conclusions
In this paper, we built an IM/DD FSO transmission system with data rate up to 10 Gbps. Based on the testbed, RSSI acquisition system was further implemented. We set up the environmental database using the parameters from NMIC in China. Then MESC was developed to repeatedly generate various channel conditions and RSSI was collected through these conditions. According to the RSSI dataset, a rolling channel prediction scheme with GRU neural network model was proposed and experiments demonstrated that the proposed scheme can predict the RSSI value and the trend of channel variation accurately. Therefore, the works will provide a helpful guideline for the design of intelligent FSO transmission system that can adapt itself to channel change.
Funding
National Natural Science Foundation of China (61775033, 61801063, 62025105); State Key Laboratory of Advanced Optical Communication Systems and Networks. (2021GZKF005); Chongqing Municipal Education Commission (KJQN201900647).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but maybe obtained from the authors upon reasonable request.
References
1. Y. Su, W. Wang, X. Hu, H. Hu, X. Huang, Y. Wang, J. Si, X. Xie, B. Han, H. Feng, Q. Hao, G. Zhu, T. Duan, and W. Zhao, “10 Gbps DPSK transmission over free-space link in the mid-infrared,” Opt. Express 26(26), 34515–34528 (2018). [CrossRef]
2. W. G. Alheadary, K.-H. Park, N. Alfaraj, Y. Guo, E. Stegenburgs, T. K. Ng, B. S. Ooi, and M. Alouini, “Free-space optical channel characterization and experimental validation in a coastal environment,” Opt. Express 26(6), 6614–6628 (2018). [CrossRef]
3. C. Abou-Rjeily, G. Kaddoum, and G. K. Karagiannidis, “Ground-to-air FSO communications: when high data rate communication meets efficient energy harvesting with simple designs,” Opt. Express 27(23), 34079–34092 (2019). [CrossRef]
4. H. Hogan, “Data demands: drive free-space optics,” Photon. Spectra. 47(2), 38–41 (2013).
5. M. A. Amirabadi and V. T. Vakili, “A new optimization problem in FSO communication system,” IEEE Commun. Lett. 22(7), 1442–1445 (2018). [CrossRef]
6. M. A. Fernandes, B. T. Brandão, P. Georgieva, P. P. Monteiro, and F. P. Guiomar, “Adaptive optical beam alignment and link protection switching for 5G-over-FSO,” Opt. Express 29(13), 20136–20149 (2021). [CrossRef]
7. P. V. Trinh, T. C. Thang, and A. T. Pham, “Mixed mmWave RF/FSO relaying systems over generalized fading channels with pointing errors,” IEEE Photonics J. 9(1), 1 (2016). [CrossRef]
8. F. J. Lopez-Martinez, G. Gomez, and J. M. Garrido-Balsells, “Physical-layer security in free-space optical communications,” IEEE Photonics J. 7(2), 1–14 (2015). [CrossRef]
9. M. A. Amirabadi and V. T. Vakili, “On the performance of a multi-user multi-hop hybrid FSO/RF communication system,” Opt. Commun. 444, 172–183 (2019). [CrossRef]
10. M. A. Esmail, H. Fathallah, and M.-S. Alouini, “An experimental study of FSO link performance in desert environment,” IEEE Commun. Lett. 20(9), 1888–1891 (2016). [CrossRef]
11. H. Kaushal and G. Kaddoum, “Optical communication in space: Challenges and mitigation techniques,” IEEE Commun. Surv. Tutorials 19(1), 57–96 (2017). [CrossRef]
12. M. A. Khalighi and M. Uysal, “Survey on free space optical communication: A communication theory perspective,” IEEE Commun. Surv. Tutorials 16(4), 2231–2258 (2014). [CrossRef]
13. A. Liang, C. Yang, C. Zhang, Y. Liu, F. Zhang, Z. Zhang, and H. Li, “Experimental study of support vector machine based nonlinear equalizer for VCSEL based optical interconnect,” Opt. Commun. 427, 641–647 (2018). [CrossRef]
14. J. Thrane, J. Wass, M. Piels, J. C. M. Diniz, R. Jones, and D. Zibar, “Machine learning techniques for optical performance monitoring from directly detected PDM-QAM signals,” J. Lightwave Technol. 35(4), 868–875 (2017). [CrossRef]
15. F. N. Khan, Y. Zhou, A. P. T. Lau, and C. Lu, “Modulation format identification in heterogeneous fiber-optic networks using artificial neural networks,” Opt. Express 20(11), 12422–12431 (2012). [CrossRef]
16. C. Rottondi, L. Barletta, A. Giusti, and M. Tornatore, “Machine-learning method for quality of transmission prediction of unestablished lightpaths,” J. Opt. Commun. Netw. 10(2), A286–A297 (2018). [CrossRef]
17. C. Zheng, S. Yu, and W. Gu, “A SVM-based processor for free-space optical communication,” in Proceedings of IEEE Conference on International Conference on Electronics Information and Emergency Communication (IEEE2015), pp. 30–33.
18. Q. Tian, Z. Li, K. Hu, L. Zhu, X. Pan, Q. Zhang, Y. Wang, F. Tian, X. Yin, and X. Xin, “Turbo-coded 16-ary OAM shift keying FSO communication system combining the CNN-based adaptive demodulator,” Opt. Express 26(21), 27849–27864 (2018). [CrossRef]
19. J. Li, M. Zhang, D. Wang, S. Wu, and Y. Zhan, “Joint atmospheric turbulence detection and adaptive demodulation technique using the CNN for the OAM-FSO communication,” Opt. Express 26(8), 10494–10508 (2018). [CrossRef]
20. M. A. Amirabadi, M. H. Kahaei, and S. A. Nezamalhosseini, “Deep learning based detection technique for FSO communication systems,” Phys. Commun. 43, 101229 (2020). [CrossRef]
21. Z. Li and X. Zhao, “BP Artificial neural network based wave front correction for sensor-less free space optics communication,” Opt. Commun. 385, 219–228 (2017). [CrossRef]
22. M. Amirabadi, M. H. Kahaei, S. A. Nezamalhosseini, and V.T. Vakili, “Deep Learning for channel estimation in FSO communication system,” Opt. Commun. 2020, vol. 459.
23. S. Lohani, N. J. Savino, and R. T. Glasser, “Free-space optical ON-OFF keying communications with deep learning,” in Frontiers in Optics (Optical Society of America, 2020), paper FTh5E.4.
24. M. Schaedler, S. Calabrò, F. Pittaì, G. Bcherer, and S. Pachnicke, “Neural network assisted geometric shaping for 800Gbit/s and 1Tbit/s optical transmission,” in Optical Fiber Communications Conference (Optical Society of America, 2020), paper M1G-1.
25. R. Fu, Z. Zhang, and L. Li, “Using LSTM and GRU neural network methods for traffic flow prediction,” in Proceedings of IEEE Conference on 31st Youth Academic Annual Conference of Chinese Association of Automation (IEEE, 2016), pp. 324–328.
26. X. Su, G. Wang, and Q. Li, “Prediction Method for Transformer State Based on GRU Network,” in Proceedings of IEEE Conference on Industrial and Commercial Power System Asia (IEEE2020), pp. 1751–1755.
27. J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” Comput. Sci. (2014).
28. M. Pavithra, K. Saruladha, and K. Sathyabama, “GRU based deep learning model for prognosis prediction of disease progression,” in International Conference on Computing Methodologies and Communication (IEEE2020), pp. 840–844.
29. M. A. Esmail, H. Fathallah, and M. Alouini, “Outdoor FSO communications under fog: attenuation modeling and performance evaluation,” IEEE Photonics J. 8(4), 1–22 (2016). [CrossRef]
30. F. Demers, H. Yanikomeroglu, and M. St-Hilaire, “A survey of opportunities for free space optics in next generation cellular networks,” in Proceedings of IEEE Conference on Annual Communication Networks and Services Research Conference (IEEE2011), pp. 210–216.
31. M. A. Esmail, H. Fathallah, and M. Alouini, “Analysis of fog effects on terrestrial Free Space optical communication links,” in Proceedings of IEEE Conference on International Conference on Communications Workshops (IEEE2016), pp. 151–156.
32. K. Li, Y. Sun, S. Li, X. Ma, and C. Zhang, “Load forecasting method for CCHP system based on deep learning strategy using LSTM-RNN,” in Proceedings of IEEE Conference on 14th Industrial Electronics and Applications (IEEE, 2019), pp. 827–831.
