Enhancing performance of next generation FSO communication systems using soft computingbased predictions Kamugisha Kazaura, Kazunori Omae, Toshiji Suzuki and Mitsuji Matsumoto Graduate School of Global Information and Telecomunicatuon Studies, Waseda University, 1011 Okuboyama, Nishi-Tomida, Honjo-Shi, Saitama 367-0035, Japan [email protected]
Edward Mutafungwa, and Timo O. Korhonen Communications Laboratory, Helsinki University of Technology, PL 3000, FIN-02015 TKK, Espoo, Finland
Tadaaki Murakami, Koichi Takahashi, Hideki Matsumoto, and Kazuhiko Wakamori Advanced Info-Communication Promotion Community, Japan
Yoshinori Arimoto National Institute of Information and Communication Technology, Japan
Abstract: The deterioration and deformation of a free-space optical beam wave-front as it propagates through the atmosphere can reduce the link availability and may introduce burst errors thus degrading the performance of the system. We investigate the suitability of utilizing soft-computing (SC) based tools for improving performance of free-space optical (FSO) communications systems. The SC based tools are used for the prediction of key parameters of a FSO communications system. Measured data collected from an experimental FSO communication system is used as training and testing data for a proposed multi-layer neural network predictor (MNNP) used to predict future parameter values. The predicted parameters are essential for reducing transmission errors by improving the antenna's accuracy of tracking data beams. This is particularly essential for periods considered to be of strong atmospheric turbulence. The parameter values predicted using the proposed tool show acceptable conformity with original measurements. ©2006 Optical Society of America OCIS codes: (010.1300) Atmospheric propagation; (010.3310) Laser beam transmission; (010.7060) Turbulence; (060.0060) Fiber optics and optical communications.
References and links 1. 2.
H. Willebrand and B. S. Ghuman, Free-Space Optics: Enabling Optical Connectivity in Today's Networks, (Sams Publishing, Indianapolis, In., 2002) T. H. Carbonneau and D. R. Wisely, “Opportunities and challenges for optical wireless: the competitive advantage of free space telecommunications links in today’s crowded marketplace,” in Wireless Technologiesand Systems: Millimeter-Wave and Optical, P. Christopher, L. Langston and G. S. Mecherle, eds, Proc. SPIE 3232, 119–128, (1998). G. Nykolak, G. Raybon, B. Mikkelsen, B. B. Brown, P. F. Szajowski, J. J. Auborn, and H. M. Presby, “160Gb/s free-space transmission link,” in Optical Wireless Communications III, E. J. Korevaar, eds, Proc. SPIE 4214, 11-13 (2001) D. Kedar and S. Arnon, “Urban optical wireless communication networks: the main challenges and possible solutions,” IEEE Commun. Mag. 42, S2-S7 (2004).
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4958
S. Bloom, E. Korevaar, J. Schuster, and H. Willebrand, “Understanding the performance of free-space optics,” J. Opt. Netw. 2, 178-200 (2003). 6. K. Kazaura, K. Omae, T. Suzuki, M. Matsumoto, E. Mutafungwa, T. Murakami, K. Takahashi, H. Matsumoto, K. Wakamori, and Y. Arimoto, “FSO Antenna with high speed tracking for improved atmospheric turbulence effects mitigation,” J. Jpn. Soc. Infrared Science and Technology (2005). 7. D. A. Montera, B. M. Welsh, M. C. Roggemann, and D. W. Ruck, “Prediction of wave-front sensor slope measurements with artificial neural networks,” Appl. Opt. 36, 675-681 (1997). 8. K. Takahashi and Y. Arimoto, “Development of optical antennas utilizing free form surface optics for the high speed laser communication systems,” in Free-Space Laser Communication Technologies XVIII , G. S. Mecherle, eds, Proc. SPIE 6105, (2006) 9. K. Gurney, An Introduction to Neural Networks, (UCL Press, London, 1997). 10. The Mathworks Inc., “Neural Networks Toolbox User Guide,” (Mathworks, Massachusets, 2005). 11. S. J. Ovaska, Computationally Intelligent Hybrid Systems: The Fusion of Soft Computing and Hard Computing (Wiley-IEEE Press, 2005) Chap. 1. 12. K. Kazaura, K. Omae, T. Suzuki, M. Matsumoto, T. Sato, K. Asatani, M. Hatori, T. Murakami, K. Takahashi, H. Matsumoto, K. Wakamori, and Y. Arimoto, “Mitigation of atmospheric effects on terrestrial FSO communication systems by using high-speed beam tracking antenna,” in Free-Space Laser Communication Technologies XVIII , G. S. Mecherle, eds, Proc. SPIE 6105, (2006)
1. Introduction Free-space optical (FSO) communication systems are being increasingly considered as an attractive option for the rapid provisioning of multi-gigabit per second links [1-3]. This is attributed to the fact that FSO systems combine the advantages of high transmission capacity enabled by optical transmission device technologies and inherent flexibility of tetherless wireless networks. As such, the deployment of an FSO link avoids the high upfront or leasing costs and time delays associated with establishing fiber-optic links, as well the stringent restrictions on the usage of the limited bandwidth in radio frequency (RF) networks. The seamless interconnection of standard singlemode fiber (SMF) and FSO links, whereby an optical signal is coupled from one media to another without any optical-electrical conversions, creates an opportunity for further cost savings. The precise coupling of a received FSO beam into an 8 – 10 µm diameter core of a SMF is enabled by the use of stateof-the-art optical antenna devices for tracking the beam's angle-of-arrival (AOA) fluctuations and steering it into the fiber core. This solution provides an effective means for maintaining acceptable link performance over long periods of time, regardless of the presence of atmospheric turbulence, building motion (due to wind sway, thermal expansion, and vibration) and so forth [4,5]. Further performance enhancements of the system are possible if some of the hard computation parameters required for controlling the beam tracking and steering devices can be predicted in advance. The aim being to reduce errors caused by temporary power dips as a result of the time delay between the evaluation of incident beam's position sensing and the resulting geometric correction of receive antenna optics. These errors become more significant in an unstable fast changing atmospheric environment with strong atmospheric turbulence  that significantly reduces the power of beam detected at the receive antenna (see Fig. 1). Almost identical predictive approaches have shown to be useful in improving image resolution of ground-based atmospheric telescopes by eliminating wavefront abberations .
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4959
Fig. 1. Beam wander and scintillation attributed to atmospheric turbulence.
Fig. 2. Experimental antenna and scintillation and optical power attenuation measurement antenna on the rooftop of Bldg. 14 Nishi Waseda Campus
2. Experimental FSO system configuration The experimental setup of the FSO communication system constituted a 1 km link providing full-duplex 2.488 Gbps data connectivity between two high-rise (9 and 10 floors) buildings located in separate campuses of Waseda University. A FSO test antenna is mounted on the rooftop of each building to enable line-of-sight communications as shown in Fig. 2. Also mounted on the rooftop of each building is an antenna used for measuring atmospheric induced intensity fluctuations (or scintillation) as well as optical power attenuation. A weather meter for measuring outdoor temperature, visibility, precipitation and fog is also installed on the rooftop of one of the buildings. In both buildings the same set of fiber transmission, monitoring, optical test measurement and equipment as well as personal computer based data acquisition systems are placed in laboratories a few floors below the rooftops for convenient access (see Fig. 3). A 1550 nm data beam incident on the compact test antenna is coupled to the inbuilding SMF (ITU-T G.652) using free form surface (FFS) mirrors in a Cassegrain arrangement and a 2 axis tip-tilt beam steering fine positioning mirror (FPM), all integrated in the antenna . The FPM has
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4960
the vital function of coupling the transmitted data beam to the SMF and steers both the transmit and received 1550nm data beam. The 980 nm beacon beam transmitted from the same antenna is used for antenna alignment and tracking purposes. At the receive antenna, this beam is isolated from the data beam and information is extracted from it for data beam tracking computations. Some of the primary specifications of the FSO communication devices used in the experiment are listed in Table 1. Error-free data reception over an FSO link is only possible if sufficient data beam power is focused on the receiver antenna. This requirement is further complicated by the challenges of adhering to an eye safety power limit and turbulence caused by channel refractive index variations attributed to variable temperature and humidity of the atmosphere [4,5]. These performance limitations can be reduced by narrow-angle beam transmission and real-time control of receiver antenna to track the movement of the incident beam. Further improvements can be attained by prediction of future values of key beam characteristics and control parameters. This advance knowledge reduces the tracking errors due to occasional time delays between the beam sensing and the completion of corrective actions of the antenna. The two parameters considered in the implementation of predictors for the test antennas are the beam angle-of-arrival (AOA) fluctuations and FPM actuator drive voltages.
Fig. 3. Photograph of experimental hardware setup. Table 1. Selected specifications of equipment used in the experiment.
E/O (directly mod. DFB laser)
Test data pattern
Boost EDFA output
Receiver optical filtrr 0.5dB BW
Rec. sensitivity (BER = 10-12)
E/O - Electrical/Optical, OOK/NRZ - On Off Keying with non return to zero, PRBS pseudo random bit sequence.
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4961
2.1 Beam angle-of-arrival (AOA) fluctuations Turbulence (manifested as scintillation and beam wander) induces power fades whereby the received signal power level might occasionally drop below the receiver sensitivity. This is also the case when there are vibrations of the devices at the antenna installation site. The common consequence of all these effects is the fluctuation of the beam AOA which directly translates to the variation of the beam power detected at the antenna thus degrading the system performance. In the experimental setup the AOA fluctuations were measured in terms of the detected electrical signal level (in Volts) when the antenna tracking was set to OFF. It has previously been noted, from comparing the power spectra, that the AOA fluctuations of the 1550 nm data beam for frequency above 100Hz are closely correlated to scintillation variations observed on the 800 nm scintillation measurement beam (see Fig. 4) . The 1550 nm test antenna and the 800 nm scintillation measurement antenna operate in same propagation path (see Fig. 2). Therefore, the AOA fluctuation parameter in terms of detected voltage variations can be considered to sufficiently representative of the position changes of the incoming data beam. The current antenna tracking and control unit (TCU) can effectively track and steer beams to the SMF for frequency less than 100 Hz therefore our main interest is to suppress the fluctuations for frequency above 100 Hz. 10 10
10 10 10 10 10 10
Intensity fluctuation caused by scintillation effects Intensity variations caused by AOA fluctuation
Fig. 4. Relationship between scintillation and rate of AOA fluctuation.
2.2 FPM actuator drive voltages Rough or coarse beam tracking by the antenna is achieved by moving the antenna using 2-axis gimbals controlled by position information that was derived from charge-coupled devices (CCD) arrays. Finer tracking is enabled when actuators move the FPM in azimuth and elevation angles depending on the incident beam's AOA. This minimizes the power coupling losses as the FPM directs a beam in to a 8 – 10 µm diameter core of a SMF with high accuracy. The actuator drive voltages are computed from the voltage outputs P1 to P4 produced the four detector elements of a quad detector (QD). The QD is able to track the incident beam's location depending on the movement of the beam spot focused on the detector. A more illustrative description of the antenna's internal structure and a schematic representing the beam tracking and control setup is depicted in Fig. 5(a) and (b) respectively. The FPM actuator drive voltages Vx and Vy trigger an actuator to move the FPM by changing its azimuth and elevation angles respectively. These voltages are directly evaluated from the QD output voltages by
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4962
Vx = Vy =
(P2 + P4 ) − (P1 + P3 )
∑i =1 Pi 4
(P1 + P2 ) − (P3 + P4 )
P i =1 i
Taking into consideration the narrow acceptance angle of SMFs (core diameter of 8.2μm and a numerical aperture of 0.13 the fiber connection incident angle is ± 7.47 degrees), the Vx and Vy error tolerance is low since the sensitivity of the FPM angular motion in azimuth and elevation angles is only about – 150 mV/deg and + 150 mV/deg respectively.
Fig. 5. The structure of test antenna (a) the cross-section view (b) configuration of internal antenna optical devices.
3. Predictor implementation 3.1 Multi-layer neural network predictor structure A soft computing based parameter predictor has the potential to complement the existing antenna's hard computing based tracking system. The resulting composite system could provide even better tracking capabilities with improved performance due to precise beam tracking even in periods of high turbulence. The soft-computing tool considered for implementing the predictors in this study are neural networks . By using data obtained from previous parameter measurements, a neural network can be trained to predict future parameter values. Figure 6 depicts the structure of a multi-layer neural network predictor (MNNP). It is composed of multiple neurons in the input and hidden layers, and a single nueron in the output layer. The MNNP structure is usually described by N – R – T nomenclature, where N, R and T is number of nuerons in the input, hidden and output layers respectively. The MNNP with N neurons in the input layer can be trained by a supervised learning algorithm using N previous and current measured parameters x(n − N − 1),..., x (n − 1), x(n ) to predict a future value
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4963
x p (n + 1) . The learning procedure is a multi-dimensional optimization problem whereby the weights and bias are varied to minimize the minimum square error (MSE) expressed by
∑ (x(n ) − x (n )) , M
where M is the number of samples in one epoch (optimization step). We employ the Levenberg-Marquardt algorithm as the learning algorithm due to its rapid convergence and accuracy in comparison to other commonly used learning algorithms . The output h j (n ) of the jth hidden neuron for a MNNP input x(n ) is given by ⎛
h j (n ) = f h ⎜ ∑ w ji x(n − 1 − i )⎟ + b j ,
Fig. 6. The structure of the multilayer neural network predictor.
where b j is the bias at the jth hidden neuron, w ji is the connection weight between the ith input and jth hidden neuron and f h is the activation function employed by the hidden layer neurons. In our study, f h is a bipolar hyperbolic tan-sigmoid function given by
f h (x ) =
2 − 1. 1 + e−x
This nonlinear function ensures the predictability of the nonlinearity in the input time series. The output layer of a MNNP for single parameter prediction has only one neuron ( T = 1 ). The neuron employs a linear summation as the activation function giving an output expressed by
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4964
y (n ) =
∑ w h (n) + b , R
where b0 is the bias at output neuron and wj is the connection weight between the jth hidden neuron and the output neuron. This output represents the next predicted value in the time series, that is, x p (n +1) = y (n ) . 3.2 MNNP integration with existing controller Several methodologies exist for integrating soft and hard computing systems to combined system with superior performance compared to systems performing individually [11, Chapt. 1]. The soft computing system considered here is the MNNP while the hard computing system is the existing tracking and control unit (TCU) in the test antenna. The possible integration configurations are (but not limited to): ∗ ∗ ∗ ∗
Parallel redundancy configurations: the MNNP is called into action when TCU loses track of the beam momentarily or fails completely. The changeover could be triggered by detected beam power falling below a certain threshold. Feedback configuration: the MNNP output is fed back into the TCU or vice versa. The aim is to provide improved transient performance of the closed loop controller implemented in the TCU. Serial configuration: the MNNP and TUC are arranged in cascade, with preceding system acting as a preprocessor for the next system. Assisted configuration: the MNNP processes the internal parameters of an active TCU prior to the TCU completing its main task.
The choice of a particular MNNP/TCU configuration will depend on several factors such as reliability, evaluation accuracy and computational delays. 4. Test results Measurement of both the training and testing data were performed in the afternoon, a period considered to be of strong turbulence causing significant errors in the link due to frequent power fluctuations. As an illustration of this case the system performance characteristics in terms of bit error rate (BER), measured received fiber-coupled power, and day temperature variations over a period of 24 hours was measured and this is shown in Fig. 7. Details on how the experimental data was collected for the FSO system performance evaluation can be found in . From the figure increased bit errors are observed at the period from around 12:00PM to about 14:00PM. This is the period considered to be of strong atmospheric turbulence because of the big temperature difference between the ground and the atmosphere leading to increase in the degree of scintillation fluctuation therefore causing burst errors. These burst errors have been attributed to either the non-linearity in the TCU or the tracking dynamic range might be insufficient in situations of strong turbulence . Alternatively, to demonstrate the difference in the degree of scintillation intensity variations as a result of scintillation can be analyzed by producing probability density functions plots by histogramming measured scintillation data for different normalized intensity values. Figure 8(b) depicts probability density function curves for intensity values of measured power fluctuations for two afternoon and evening time series data shown in Fig. 8(a). Qualitatively the wider the histogram (see Fig. ((b)) the greater the degree of received signal fluctuations due to scintillation. The magnitude of scintillation is a strong function of the time of day and ambient temperature. This confirms that the degree of scintillation is stronger during the afternoon than any other times of the day. The time series data shown in Fig. 8(a) and the AOA fluctuation measured data (see Fig. 9(a)) is recorded after every 5 minutes in the current experimental setup. The 5 minute periods
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4965
are split into blocks of 3 seconds wherein the sampling rate is 10 kHz. A 3 second (3×104 samples) AOA fluctuation data block (shown in Fig. 9(a)) was used in the training phase while the subsequent 3 second block was used for the test phase. Although the FPM tracking speed is set at 1 KHz, the actuator drive voltage data is only recorded once per second. This Vx and Vy data is bundled in to 305 second blocks (305 samples), whereby a single block was used for the training phase (see Fig. 9(b)) and subsequent block for the testing phase. 1.0E-00 1.0E-01
Temperature (ο C)
Bit Error Rate
1.0E-04 Bit error rate Received power Temperature
1.0E-06 1.0E-07 1.0E-08
1.0E-11 Error free 0:00
12:00; Temp. 32οC
2 1 0
Monitor output (V)
Time (msec) 4 ο
18:00; Temp. 23 C
1.4 1.2 1 0.8 0.6
Monitor output (V)
Fig. 7. Bit error rate and fiber received power characteristics
2000 Time (msec)
1.5 2 2.5 Normalized Intensity
Fig. 8. Measured beam intensity fluctuation (a) time series data (b) probability density function of times series data.
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4966
5 Okubo campus
3 Okubo campus 2
4 Drive voltage (V)
Detected signal level (V)
1 0 -1 -2
150 200 Time (sec)
Fig. 9. Predictor training data (a) AOA fluctuations training data and (b) FPM actuator drive voltages training data.
After repeated trials a 5 – 20 – 1 MNNP was selected for prediction of both the AOA fluctuation and FPM actuator drive voltages. The respective prediction errors for test data are shown in Fig. 10(a) and 10(b). Mean prediction errors for Vx and Vy are – 35.1 µV and – 0.4 µV respectively. This implies that typical FPM angular motion errors caused by the predictor are only – 4 µrad (azimuth) and – 0.05 µrad (elevation) obtained from the average prediction errors of the FPM actuator voltages and using the values of the sensitivity of the FPM angular motion in azimuth and elevation quoted in section 2.2. A more clear view of the accuracy of the MNNP is provided by regression analysis between the predicted values and the test data. The analysis produces three results: the correlation-coefficient β, the y-intercept
a = xp
and the slope r =
of the best linear regression between the measured and
Prediction error (μV)
predicted values. A perfect prediction of the parameter test values would return the values β = 1.0, α = 0, and r = 1.0 . The results in Table 2 indicate the high accuracy of all prediction tests performed.
0 -0.25 -0.5 -0.75 -1 0
Prediction error (mV)
Prediction error (V)
1 0 -1 -2 0
Fig. 10. Errors in the test prediction (a) of detected voltage variations due to AOA fluctuations and (b) of FPM actuator drive voltages.
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4967
Table 2. Regression analysis results for different predictions.
FPM actuator drive voltage Vx
8.44 x 10-7
FPM actuator drive voltage Vy
-9.64 x 10
5. Conclusions and practical considerations The idea of an FSO antenna that can be trained directly with previous measurements may be realized using MNNPs. By employing the future parameter values, the errors that occur during the periods of high turbulence may be reduced. This is possible by enabling the antenna control system to utilize predicted values when there are excessive hard computing delays or uncertainties in the original tracking devices. Furthermore, the MNNP implementations are low cost and could be engineered to provide predictions with minimal time delay. Future work involves designing the predictor circuit and making a hardware implementation into the antenna tracking and control mechanism. Experiments on the performance improvements after intergrating the MNNP to the antenna will be conducted. Acknowledgments This work is supported by a grant from the National Institute of Information and Communications Technology (NiCT) of Japan.
(C) 2006 OSA
12 June 2006 / Vol. 14, No. 12 / OPTICS EXPRESS 4968