Comments: download pdf
file .
Ref.: Proc. OCEANS MTS/IEEE ' 2007,
(ISBN:), p., Vancouver, Canada, 2007. (to appear)
Abstract : This
paper presents experimental results for equalization of single-carrier
transmissions using multichannel RLS lattice filters under fixed-point
arithmetic. The performance of the lattice algorithm is illustrated and
compared with other RLS algorithms using real data from the MREA'04 and
MAKAI'05 ocean experiments. Recently,
the implementation of signal processing algorithms in dedicated
hardware has gained new momentum due to the availability of (1)
large-capacity FPGAs that incorporate hardware multipliers and other
useful features for mathematical computation, and (2) tools that can
automatically generate reasonably-efficient HDL code from block
diagrams or other high-level descriptions. Such power-efficent and
potentially highly-parallel solutions are certainly appealing for
equalization of underwater channels, where the computational complexity
increases with the square of the number of coefficients, which is often
large (long impulse responses, multiple receivers). However, the input
covariance matrix tends to be ill-conditioned when fractional sampling
or multiple receivers are used, especially with high filter orders, and
in turn this leads to high dynamical ranges in the internal variables
of several adaptive filtering algorithms, including plain RLS.
Accomodating such numerical excursions in fixed-point arithmetic
requires large word lengths, and hence more FPGA real estate and power
consumption.
A new multichannel lattice algorithm for RLS filtering was
proposed in [1]. The algorithm is based on a modular decomposition
approach that enables unequal channel orders to be specified and leads
to a filter that is structured as an array of interconnected scalar
units which operate similarly to those found in conventional
single-channel lattice filters. This avoids cumbersome matrix
operations that are difficult to implement in hardware and not easily
parallelizable. For a least-squares problem with L channels and M
adjustable coefficients per channel the filter complexity is
proportional to ML2, rather than (ML)2 as in plain RLS. Each scalar unit performs
an input-output mapping independently of all the remaining ones, so
that a fully-parallel implementation may be obtained by adding pipeline
registers between units.
By exploiting a known link with Kalman filtering theory, the
update recursions for individual units are written in array form, i.e.,
forward or backward (2-by-2) prearrays are built from each unit's
internal variables, a Givens rotation is applied to create a postarray
where a specific element is zeroed-out, and time-updated variables are
read from other entries in the array. CORDIC algorithms enable Givens
rotations to be iteratively implemented in hardware using only simple
add/shift operations. The array formulation is known to exhibit
excellent numerical properties in the single-channel case, and these
were found to carry over to the proposed multichannel lattice. In
simple simulated scenarios the array-based lattice exhibited none of
the numerical instability problems that affect other fast RLS
algorithms even for 10-bit fixed-point precision. Its behavior was
found to be consistent in the sense that the output mean-square error
(MSE) progressively increased as the numerical precision decreased.
Moreover, for any given precision its MSE was lower than that obtained
with other multichannel lattice and plain RLS algorithms.
This work examines the
fixed-point performance of the array-based multichannel lattice under
more challenging conditions, using data from two underwater
communication experiments:
The MREA'04 (Maritime Rapid Environmental Assessment) sea trial
was conducted in the continental shelf off the west coast of Portugal
in April 2004, in an area to the north of the Setúbal Canyon.
The acoustic source was suspended at a depth of about 60 m, and the
receiver was an 8-element drifting array with hydrophones placed at
depths 10, 15, 55, 60, 65, 70, 75, 80 m. During a period of
approximately 90 minutes modulated data were transmitted over a range
of about 2 Km, using a carrier frequency of 3600 Hz, symbol rates of
200 or 400 baud, and both 2-PSK and 4-PSK constellations. The
transmitter was being towed throughout much of the experiment at speeds
of up to 2 m/s, thus inducing significant Doppler scaling in received
waveforms.
The MAKAI'05 experiment took place off the island of Kauai,
Hawaii, in September/October 2005, in an area that is part of the
Pacific Missile Range Facility (PMRF). MAKAI'05 was specifically
planned to support the High-Frequency initiative, whose goal is to gain
a better understanding of acoustic propagation at frequencies on the
order of tens of kHz. A large number of teams from various countries
and institutions were involved, each focused on a specific set of
objectives related to its equipment and scientific goals. The data
examined in this paper pertains to an experiment designed by the
University of Algarve, Portugal, in which 2-PSK data were transmitted
at 2000 baud (3 kHz bandwidth) around a carrier frequency of 10 kHz.
Both moored and towed sources were used, and source-receiver ranges
were on the order of 1 to 4 Km. The drifting receiver array was similar
to the one used in MREA'04, but only 5 hydrophones were functional.
Both linear (FSE) and
decision-feeback (DFE) equalizer architectures are considered in this
work. Contrary to the usual transversal DFE, where the outputs of the
feedforward and feedback filters can be individually accessed, all
channels (fractionally-sampled multichannel acoustic data and
previously-decoded symbols) are merged in a lattice DFE. This excludes
some widely-used algorithms for carrier and symbol synchronization in
DFE based on the MMSE criterion, and a few alternative options,
MMSE-based or not, are discussed.
As in [1], it is found
that the differences in performance between RLS variants are small
under floating-point arithmetic. However, only the array-based lattice
provides acceptable performance for 12-bit fixed-point precision. Plain
RLS performs very poorly even for 16-bit precision.
[1] J. Gomes, V. Barroso, "Array-Based QR-RLS Multichannel Lattice Filtering," Submitted to IEEE Transactions on Signal Processing, April 2007.
ACKNOWLEDGMENT: this work was partially supported by
FCT projects NUACE - POSI/CPS/47824/2002 and RADAR -
POCTI/CTA/47719/2002.
The following copyright notice
applies to papers that have appeared (or will appear) in IEEE
publications:
“©1992-2007 IEEE. Personal use of this material is permitted.
However, permission to reprint/republish this material for
advertising or promotional purposes or for creating new collective
works for resale or redistribution to servers or lists, or
to reuse any copyrighted component of this work in other works must be
obtained from the IEEE.”