matrices, etc. Additionally Beamforming may be applied on the
fly. Field tests in private homes show the significantly increased
data throughput and link reliability of the MIMO PLC system.
Keywords- Power line communication; MIMO
I. INTRODUCTION
The application of Multiple Input Multiple Output (MIMO) methods to inhome power line communications (PLC) promises significantly higher data rates and better coverage compared to today’s Single Input Single Output (SISO) systems [1-4]. The HomePlug Powerline Alliance considers MIMO technique as the key element of next generation PLC modems [5]. A feasibility study was undertaken by Sony to verify the advantages of MIMO PLC compared to SISO transmissions.
MAM
TX, 2 paths
RX, 4 paths
MIMO probe
GUI
Figure 1. Hardware setup of the MIMO PLC feasibility study
Fig. 1 shows the hardware setup of our study. The PC on the left-hand side operates as the transmitter (Tx) where two coaxial cables are connected to the MIMO PLC triangle-style probe [6]. The Tx probe is connected to the outlet of the MAM (MIMO Artificial Mains, details are presented in [7]) which consists of the three boxes Front End A, Attenuator Unit and Front End B. To the top of the MAM units are various filter and attenuator elements connected causing a high diversity MIMO PLC channel. From the receiving star-style probe [6] four coaxial cables lead to the receiving MIMO PLC modem
which is implemented in the 2nd PC on the right of Fig. 1. The number of paths connected to the system is an indication to the dimensions of the MIMO matrix: 2x4.
The MIMO PLC system is able to stream high definition (HD) video or UDP data via PLC. The transmission uses the frequency range from 4 MHz to 30 MHz. The system operates with orthogonal frequency division multiplexing (OFDM) and adaptive modulation.
This paper gives first a detailed overview of the system architecture implemented in the MIMO PLC feasibility study (Section II). Next, channel estimation (Section III), MIMO equalization (Section IV) and Beamforming (Section V) are discussed in more detail. Section VI introduces the results of a field test in a private home. Here, the performance of the system is compared when operated in SISO and MIMO mode. Several screen shots visualize the performance of MIMO PLC.
II. SYSTEM ARCHITECTURE
As described in Section I the feasibility study is implemented on two PCs and the MAM. Each of the PCs embeds a PCI card including a Xilinx SX 95 and LT 330 field-programmable gate array (FPGA). The operation system is Linux. The software driver is a standard Ethernet driver extended with the PLC functions like access to the FPGAs and application requests. The architecture of the software stack is similar to the smart (or dynamic) notching system described in [6]. The embedded AFEs (Analog Front Ends) in the system are modified evaluation boards from Analog Devices AD9867.
TABLE I. BASIC PHYSICAL LAYER (PHY) PARAMTERS
MIMO setup 2 Tx and 4 Rx
FFT points 2048
Sampling frequency (MHz) 80
Frequency band (MHz) 4-30
Number of active subcarriers (4-30MHz) 1296
Carrier spacing (kHz) 19.53
Symbol length (µs) 51.2
Guard interval (µs) 3.2 (1/16)
Modulation (per subcarrier) QPSK, 16-, 64-, 256-,
1024-QAM
Error correction RS (204,188), Viterbi
(coderate ½ or ¾ )
Maximum transmission speed, PHY (Mbit/s) 506
2011 IEEE International Symposium on Power Line Communications and Its Applications
Each of the PCs can be configured as transmitter (Tx) as well as receiver (Rx). In Tx mode two of the 10 bit digital-to-analog converters (DAC) and in Rx mode all four analog-to-digital converters (ADC) are under operation. The implemented applications are the transmissions of HD video or UDP test data. A third PC running Matlab allows to control the system and to visualize the configuration status. The target of the project is to verify the advantages of MIMO PLC compared to SISO transmissions. The expected gain from the simulation results published in [1-2] are verified in a hardware implementation. Rapid design was a design criterion in the development phase. Fig. 2 shows the block diagram of the MIMO PLC system.
Figure 2. Block diagram of MIMO PLC sytsem
The systems forward error correction (FEC) and OFDM parameters are identical to the SISO system presented in [6]. The PLC system is extended with the MIMO functions of a 2
nd
transmitter, 3 additional receivers and the MIMO processing. The key features are summarized in Table I. Fig. 2 shows the Tx chain from left to right in the top row. The video data is streamed via DMA to the FPGAs. The FEC implemented is a classical approach of Reed Solomon and Viterbi codecs and in addition the interleavers. MIMO functions start at the demultiplexer (S/P) when the single data stream is split into the 2 logical MIMO streams. Each bit stream is modulated by the quadrature amplitude modulation (QAM) block. The system is using a carrier individual adaptive QAM ranging from 4-QAM to 1024-QAM for each of the 2 MIMO streams. A simplified power allocation (PA) is applied on the 2 MIMO streams. If one stream carries no information the power is allocated to the other stream resulting in 3 dB SNR gain of this path. Next, Beamforming is applied at the 2x2 V-matrix multiplication in the precoding block (see Section V). 4 Training Symbols are inserted before the OFDM block which is implemented by a 2048 point inverse fast Fourier transform (IFFT). The Tx filter comprises the mixer to shift the base-band signal to the useable frequency range and the line interface filter including the DAC. A triangle-style or transversal coupler (not shown in Fig. 2) feeds the Tx signals onto the mains. The receiver chain of the MIMO PLC system is depicted in the lower row of Fig. 2 from right to left. A star-style (longitudinal) coupler receives three differential mode (DM) signals and the common mode (CM)
signal from the mains. Four AFEs capture the signals. The Rx filter block is responsible for three functions: first it includes shifting the sampled data into the base band, second Rx time synchronization to set the trigger to the median result out of the four streams and third the automatic gain control (AGC) to adjust the four AFEs individually. The OFDM demodulation comprises a fast Fourier transform and transforms the OFDM symbols from time into frequency domain. In the next block the data burst is split into training symbols and data. The training symbols are processed in the channel estimation (see Section III). The MIMO detection is based on the pseudo inverse of the channel matrix and equalizes the four received streams into the two logical MIMO streams (see Section IV). The SNR estimation includes an equalizer, a normalization utilizing the known training symbols and a variance calculation. The block codebook search determines the precoding matrices for each subcarrier (see Section V). For simplicity the training symbols are not precoded in this implementation. A MIMO PLC product should beamform its full preamble to utilize maximum MIMO gain. Receiver Beamforming is done before the two streams are QAM demodulated. A combiner (P/S) multiplexes the data which have been split at transmitter back into a single stream. The FEC removes the error protection added at the transmitter. The SNR estimation results define the choice of the QAM constellations for adaptive modulation in the PLC driver. The constellations and precoding information are adapted at the transmitter and receiver simultaneously.
III. MIMO PLC CHANNEL ESTIMATION
To estimate the channel the training symbol t is transmitted four times in a normal (“+”) or inverted (“-“) way on each transmitter to achieve orthogonality. To enable an estimation of each path individually Tx 1 sends the symbol in the way “+ + - -“ and simultaneously Tx 2 sends the symbol in the way “+ - + -“. The received symbols for the four time instances (denoted by the upper index) on receive port i are
( )ththr iii 21
1 += , ( )ththr iii 21
2 −= ,
( )ththr iii 21
3 +−= , ( )ththr iii 21
4 −−= (1)
with ijh the channel coefficient from transmit port j to
receive port i. Due to the spatial orthogonality of the training symbols the channel coefficient from each transmit port are estimated for each receive port
( ) ( ) ( ) ( )
t
rrrrh iiii
i4
4321
1
−−+= ,
( ) ( ) ( ) ( )
t
rrrrh iiii
i4
4321
2
−+−= . (2)
The absolute value of the training symbol t is equal to 1. Thus, the division in (2) simplifies to a multiplication by the conjugate complex of t. The channel estimation is performed for each subcarrier separately.
55
The magnitude of the received training symbols after AGC and ADC are shown in Fig. 3. Each of the 16 graphs has the frequency range from 0 Hz up to 40 MHz in the x-axis and the magnitude of the received training symbol in the y-axis in dB. To prove reproducibility of the results always five traces in different colors are plotted. The 16 graphs are aligned to have Rx port 1 in the left column up to Rx port 4 in the right column. The 1
st received symbol is in top row, the 4
th symbol
in the bottom row. Due to overlapping or elimination of the two transmitted signals the received signal levels vary from each received signal and receive port by more than 20 dB.
Receiv
ed
tra
inin
g s
ym
bols
RX1 RX2 RX3 RX4
1st
2nd
3rd
4th
Figure 3. Received training symbols
Fig. 4 presents the channel estimation calculated from received training symbols. The 2x4 MIMO channel has eight individual paths. Again, each sub-plot represents the transfer function over frequency. The scaling of the y-axis is arbitrary selected from the output of the channel estimation in hardware. The top row illustrates the four channels from transmit port 1, the bottom row from transmit port 2. From left to right are the four Rx ports.
2x4
Ch
an
ne
l T
ran
sfe
r fu
nctio
ns
TX1
TX2
Figure 4. Channel estimation calculated from received training symbols
The channel estimation result of Fig. 4 shows that there is a high diversity between the eight paths. The fading characteristics – typical for PLC channels – are different for each path. If due to the multi-path PLC channel a transmitted carrier is eliminated on one path the carrier is available on any other path. MIMO technology eliminates deep fadings. This shows how MIMO PLC supersedes SISO PLC systems.
IV. MIMO DETECTION AND EQUALIZATION
The performance of beamfoming (see next section) is independent of the detection algorithm. Thus, the most simple equalizer is used which is zero-forcing (ZF) detection. Zero-forcing detection requires the calculation of the pseudo inverse H
P of the channel matrix H [2]. An efficient and numerically
stable implementation of the pseudo inverse calculation is achieved by a QR decomposition of the channel matrix H = QR where Q is a unitary matrix and R is an upper triangular matrix. The pseudo inverse and equalizer matrix is then calculated as
( ) HHHp QRHHHHW 11 −−=== . (3)
The implemented QR decomposition is based on iterative unitary rotations to transform H into the upper triangular matrix R. The unitary transformations are implemented by the cordic algorithm [8]. Fig. 5 shows the block diagram of the equalization process. Based on the estimated channel matrix H, the QR decomposition is calculated. In a next step, the inverse of R is calculated. The upper triangular structure of R simplifies the inversion. Q is represented by several rotation angles. The application of the rotation angles on the received data and the multiplication by R
-1 perform the equalization.
Figure 5. Zero-forcing detection using QR decomposition
V. BEAMFORMING BASED ON A CODEBOOK
A comparison of different MIMO schemes showed that Beamforming offers the best performance [1]. Beamforming requires channel state information at the transmitter. The receiver has to feedback the information of the precoding matrix to the transmitter. For PLC this feedback path already exists due to the adaptive modulation and the feedback of constellations. An appropriate quantization limits the feedback overhead to the same order of magnitude as the feedback of the constellation information [2]. The optimum Beamforming matrix V is obtained from a singular value decomposition (SVD) of the channel matrix H [2]. V is a unitary 2x2 matrix. Thus, the two column vectors are orthogonal and the norm of each column vector is equal to one. These properties allow to describe the precoding matrix V by two independent parameters: the real value v1 and the angle φ2
−=
12
*21
vv
vvV with 22
12 1φj
evv −= . (4)
The feedback of the precoding matrix is based on a codebook. Both, the receiver and transmitter have a predefined set (look-up table) of precoding matrices. The codebook contains the two parameters v1 and φ2 for each stored precoding matrix. The receiver determines the best fitting precoding matrix and feeds back the index in the codebook to the transmitter. Here, a feedback of 7 bit per carrier is implemented, i.e. the codebook contains 128 entries. The codebook is designed to represent uniformly distributed precoding matrices [2].
56
The precoding matrix influences the detection matrix. The new detection matrix comprises the hermitian transpose of the precoding matrix and the pseudo inverse of the channel matrix
pH HVW = . (5)
Based on the detection matrix the equivalent SNR after detection for ZF is calculated as
2
1
1
1SNR
wρ= and
2
2
2
1
wSNR ρ= (6)
with ρ the ratio of transmit power to noise power and iw
the norm of the i th row of the detection matrix W.
Different precoding matrices influence the SNR after detection according to (5) and (6). The contour lines in Fig. 6 show the SNR of the first MIMO path for different precoding matrices described by the two parameters v1 and φ2. The black horizontal lines limit the plane to +/- π. The codebook entries are marked by the black squares. The filled black square shows the codebook entry closest to the optimum precoding matrix (black diamond symbol). As a reference spatial multiplexing without precoding (v1 = 1 and and φ2 = 0) is also indicated.
25dB
v1
φ2
0 0.2 0.4 0.6 0.8 1
−6
−4
−2
0
2
4
6
26
28
30
32
34
36
38
40
42
44
46
48
SNR1(v
1, φ
2, H)
codebook entries
codebook entries (φ2 +/− 2π)
optimum precoding
found codebook index
no precoding
Figure 6. SNR of the 1st MIMO path depending on precoding, codebook
entries
To avoid a brute-force search through all codebook entries, an efficient search algorithm is implemented. Based on a coarse quantization of the optimum precoding matrix only the codebook entries near the quantized precoding matrix are checked through the search.
Fig. 7 shows ║wi║2 which is proportional to the SNR
according to (6) of the two logical MIMO streams. In each plot the horizontal axis represents the frequency range from 4 MHz to 30 MHz and the vertical axis the signal level in dB. The SNR of each stream when spatial multiplexing is used is plotted with the green lines. When Eigenbeamforming is applied usually the SNR of one stream benefits significantly
and the other one is slightly mitigated. This effect is plotted with the cyan line for SNR1 and the black line for SNR2.
The MIMO PLC feasibility study allows monitoring of the received symbols with and without Beamforming in the constellation diagrams. The constellation diagrams are shown in Fig. 8. Each of the four graphs represents the complex plane in the frequency domain. The graphs in the top row are recorded without Beamforming, the graphs in the bottom row with Beamforming. The left column shows MIMO stream 1 and the right column stream 2. Each dot in Fig. 8 represents an OFDM carrier in one symbol. Here several successive data symbols are plotted into one chart using various colors. Due to varying noise on each symbol the dots do not fully overlap. Constant 4-QAM was transmitted on all carriers. When comparing the constellations when Beamforming is applied and not it is visible how stream 1 benefits from the better SNR and stream 2 shows increased noise.
VI. MIMO PLC FIELD TEST IN PRIVATE HOMES
The feasibility study has been tested in private homes to verify the results out of simulations under real conditions. For comparison the system is operated in SISO and in MIMO mode. Transmission in SISO mode is performed with one AFE, only, where MIMO transmitting utilizes two AFEs. This results in the double injected feeding power for MIMO transmission. The results presented in this paper are not influenced by the increase of transmission energy because the receiver’s AGC settings were operating anywhere in the middle of their dynamic range. Therefore the additional feeding power was compensated at receivers AGC. The higher feeding power
57
would have influenced throughput results if the receiver would operate at maximum gain settings.
Equalizer matrices, QAM
Figure 9. The application was watching an HD soccer match
Fig. 9 shows the application running on the MIMO PLC system. The VAIO on the right side renders the MPEG 2 coded HD video data. The monitor displays the GUI of the MIMO PLC control and monitor system. Target of the tests was to compare throughput rates of SISO and MIMO transmissions. Therefore a transmission with the optimal constellations is setup where the HD video never showed blocky artefacts or the Reed Solomon decoder failed to correct errors. Selected constellations, data throughput rates, AGC settings, Viterbi and RS error rates were monitored.
- 11 -
Triangle Style
Probe
Figure 10. Photo of the transmitter in a private home
The transmitter is shown in Fig. 10 feeding two signals into the mains. Two coaxial cables from the PC to the PLC triangle probe are visible. Fig. 11 shows the receiver connected to the mains. The star coupler is tightly connected to a large copper plate which provides a large capacity to the ground. This enables low impedance grounding for HF signals. The
measurement results do not change when a human is touching one of the devices.
Figure 11. Photo of the receiver in a private home
SNR constellations1024-QAM256-QAM
64-QAM16-QAM
QPSK
RS failed = 0: FEC corrects all transmission errors.
143 Mbit/s
Figure 12. GUI showing the results of a SISO transmission
Fig. 12 and Fig. 13 show a screenshot of the GUI controlling and monitoring the PLC system when communicating in SISO and MIMO mode, respectively. Beneath the control buttons the screenshot includes four subplots presenting (from top to bottom) SNR and selected constellations, bit error rates, throughput rates and AGC settings. The top subplot represents the frequency range from 4 MHz to 30 MHz in horizontal axis and the SNR in vertical axis. First, SISO transmission in Fig. 12 is considered. The thin line shows the SNR measured by the system. The bold line corresponds to the constellations selected by the PLC driver at this frequency. Four different levels or steps of constellations are visible: the lowest one is 4-QAM, next 16-QAM, 64-QAM and 256-QAM. The maximum constellation of 1024-QAM is not selected via this channel. All of the three lower subplots represent the time in x-axis. One minute history is shown scrolling from right to the left. A small error rate of 0.0035 at Viterbi and 17 at RS is corrected when the screen was shot. Both numbers indicate the number of bits corrected by FEC within any time unit. Important is that the RS failed information never exceeds zero. This would indicate an uncorrectable error. The throughput rate on raw phy is 143 Mbps (red line of the 3
rd subplot from the top). AGC
settings show a gain of 14 dB, 11 dB, 9 dB and 7 dB for the
Counterpoise
for CM
reception
Star style
probe
58
individual receivers. The channel connected here was relatively low attenuated.
Fig. 13 shows the GUI of the MIMO PLC system connected to the same pair of outlets as when Fig. 12 was recorded but transmission operated in MIMO mode including Beamforming. The main changes are visible in the top plot where two SNR measurements (thin lines) and two constellation maps (the bold cyan and black line) are displayed. At many frequencies OFDM carriers of both MIMO streams carry 64-QAM or even higher constellations. The 2
nd subplot
shows that all channel errors were corrected by the FEC, as well. The raw phy rate increased to 315 Mbps which is more than twice the SISO rate. AGC gains were reduced by 2 dB to 3 dB compared to SISO transmission because of the feeding of two transmitters.
VII. CONCLUSIONS
A MIMO PLC real-time system is implemented to prove the feasibility of MIMO for PLC. The system supports two transmit and four receive ports. A field test performed in a private home proves the increase of MIMO PLC throughput rates compared to SISO transmissions. The example documented in this paper shows an improvement of factor 2.2. An HD video is transmitted error free via SISO and MIMO PLC. Next steps are to collect a large statistics in several buildings.
REFERENCES
[1] L. Stadelmeier, D. Schneider, D. Schill, A. Schwager and J. Speidel, “MIMO for Inhome Power Line Communications“, International Conference on Source and Channel Coding (SCC), 2008.
[2] D. Schneider, L. Stadelmeier, J. Speidel and D. Schill, “Precoded Spatial Multiplexing MIMO for Inhome Power Line Communications“, Global Telecommunications Conference (Globecom), 2008.
[3] R. Hashmat, P. Pagani, A. Zeddam and T. Chonavel, “MIMO communications for inhome PLC networks: Measurements and results up to 100 MHz“, International Symposium on Power Line Communications and Its Applications (ISPLC), 2010.
[4] A. Canova, N. Benvenuto and P. Bisaglia, “Receivers for MIMO-PLC channels: Throughput comparison“, International Symposium on Power Line Communications and Its Applications (ISPLC), 2010.
[6] A. Schwager, “Powerline Communications: Significant Technologies to become Ready for Integration”, Ph.D. dissertation, Universität Duisburg-Essen, 2010.
[7] A. Schwager, D. Schneider, W. Bäschlin, A. Dilly and J. Speidel, “MIMO PLC: Theory, Measurements and System Setup”, International Symposium on Power Line Communications and Its Applications (ISPLC), 2011.
[8] C.M. Rader, “VLSI Systolic Arrays for Adaptive Nulling”, IEEE Signal Processing Magazine, vol. 13, pp. 29-49, 1996.
SNR (logical) MIMO path 1 constellations path 11024-256-
64-16-
4-
RS failed = 0: FEC corrects all transmission errors.
QAMSNR (logical) MIMO path 2
constellations path 2
315 Mbit/s
Figure 13. GUI showing the results of MIMO transmission
