U.S. patent application number 14/714536 was filed with the patent office on 2015-09-03 for adaptive demodulation method and apparatus using an artificial neural network to improve data recovery in high speed channels.
The applicant listed for this patent is Oluwatobi Olabiyi, Dhadesugoor Vaman. Invention is credited to Oluwatobi Olabiyi, Dhadesugoor Vaman.
Application Number | 20150249554 14/714536 |
Document ID | / |
Family ID | 54007260 |
Filed Date | 2015-09-03 |
United States Patent
Application |
20150249554 |
Kind Code |
A1 |
Vaman; Dhadesugoor ; et
al. |
September 3, 2015 |
ADAPTIVE DEMODULATION METHOD AND APPARATUS USING AN ARTIFICIAL
NEURAL NETWORK TO IMPROVE DATA RECOVERY IN HIGH SPEED CHANNELS
Abstract
A neural network demodulator is used within a receiver to
provide Inter Symbol Interference (ISI) channel equalization and to
correct for I/Q/phase imbalance. The neural network is trained with
a single integrated training step to simultaneously handle the
channel impairments of ISI equalization and I/Q phase imbalance as
opposed to prior art methods of separately addressing each channel
impairment in sequence.
Inventors: |
Vaman; Dhadesugoor; (Spring,
TX) ; Olabiyi; Oluwatobi; (Herndon, VA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Vaman; Dhadesugoor
Olabiyi; Oluwatobi |
Spring
Herndon |
TX
VA |
US
US |
|
|
Family ID: |
54007260 |
Appl. No.: |
14/714536 |
Filed: |
May 18, 2015 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
14312072 |
Jun 23, 2014 |
9036745 |
|
|
14714536 |
|
|
|
|
61837742 |
Jun 21, 2013 |
|
|
|
Current U.S.
Class: |
375/232 |
Current CPC
Class: |
H04L 25/03006 20130101;
H04L 25/03159 20130101; H04L 2025/03636 20130101; H04L 25/03821
20130101; H04L 25/03019 20130101; H04L 25/03165 20130101 |
International
Class: |
H04L 25/03 20060101
H04L025/03 |
Claims
1. A method for use in a communication system, the communication
system including a transmitter and a receiver, to provide ISI
channel equalization and to correct I/Q phase imbalance,
comprising, providing a demodulator as part of the receiver, the
demodulator including a trainable neural network, and training the
neural network in a single integrated training step to
simultaneously enable the demodulator to provide ISI channel
equalization and to correct I/Q phase imbalance.
2. A method in accordance with claim 1 wherein the neural network
includes two layers, the method further comprising the step of
implementing a transfer function between the two layers of the
neural network to ensure convergence of the neural network training
step.
3. A method in accordance with claim 2 wherein transmitter data
symbols and RF carrier signals are sent over the communication
channel from the transmitter to the receiver with the I/Q imbalance
resulting from carrier phase misalignment between the transmitter
and receiver, said neural network training step utilizing
information sent over the communication channel to determine
coefficients needed by the neural network to correct said carrier
phase misalignment.
4. A method in accordance with claim 3 wherein transmitter data
symbols and RF carrier signals are sent over the communication
channel from the transmitter to the receiver with ISI resulting by
the change of bandwidth of the carrier signal frequency causing
interference between adjacent transmitter data symbols, said neural
network training step utilizing information sent over the
communication channel to determine coefficients needed by the
neural network to prevent the interference between transmitter data
symbols.
5. A method in accordance with claim 4 wherein the neural network
training process determines coefficients W1, b1 W2 and b2 such that
the mean squared error between a transmitted symbol and a symbol
received at the output of the demodulator is minimized.
6. A communications system, comprising, a transmitter for sending
modulated data symbols and carrier signals over the communications
channel, a receiver including a demodulator, and a trainable neural
network included as part of the demodulator, which when trained in
a single integrated training step allows the demodulator to
simultaneously provide ISI channel equalization and correct I/Q
phase imbalance.
7. A communications system in accordance with claim 6, wherein the
neural network includes two layers, one hidden layer and one output
layer, the hidden layer applying a first set coefficients to data
symbols from the receiver to create a first set of fixed point
words, and processing the first set of fixed point words with a
transfer function.
8. A communications system in accordance with claim 7, wherein the
first set of fixed point words are sent to the output layer after
processing by the transfer function and applied to a second set of
coefficients.
9. A communications system in accordance with claim 8, wherein the
neural network is trained based on information sent over the
communication channel to generate the first and second set of
coefficients.
10. A communications system in accordance with claim 9, wherein the
first and second set of coefficients ensure that the mean squared
error between a transmitted data symbol and a symbol at the output
of the demodulator is minimized.
Description
RELATED APPLICATIONS
[0001] This application is a continuation-in-part of U.S. patent
application Ser. No. 14/312,072, filed Jun. 23, 2014, hereby
incorporated herein by reference, and U.S. Provisional Patent
Application No. 61/837,742, filed Jun. 21, 2013.
FIELD OF THE INVENTION
[0002] The present invention relates to a neural network based
integrated demodulator that mitigates channel impairments, ISI and
I/Q channel leakages, and minimizes the impact on the overall
performance of the system. In this process, the inventive
demodulator improves the performance for data recovery when
operating at high data rates over the transmission channel.
BACKGROUND OF THE INVENTION
[0003] High-Speed communications systems typically use a wide band
channel where the transmission is achieved using Radio Frequency
(RF) carriers. RF transmission as it exists today either uses
closely spaced narrow band multiple carriers or a small number of
carriers containing baseband modulated signals.
[0004] An example of a closely spaced narrow band multiple carrier
system is the Orthogonal Frequency Division Multiplexing (OFDM)
system, which uses a large number of RF carriers with each carrier
carrying two base band modulation signals (I/Q). Since OFDM uses
orthogonal carriers, the transmission does not suffer any
Inter-Frequency-Interference (IFI). Also, the data processing at
the receiver uses a simple Fast Fourier Transform (FFT) technique.
Due to the orthogonality of the RF carriers, an OFDM transmission
system is more robust to Inter-Symbol-Interference (ISI). However,
when ISI exists, the system requires the use of a Cyclic Prefix as
an overhead and a channel equalizer to handle ISI.
[0005] An example of a small number of carriers containing baseband
modulated signals is the "Kelquan" system based on the teachings
presented in U.S. Patent Nos. 5,956,372 and 8,233,564 where closely
spaced non-orthogonal frequencies are used to create baseband
modulated signals which are carried in a small number of RF
carriers as I/Q channels over a wideband bandwidth. In this system,
the data is recovered optimally after the IFI suppression using a
Neural Network Matched Filter. This system requires no overhead,
but needs a robust equalizer to handle ISI.
[0006] In both of the above scenarios, the performance of
high-speed digital transmission suffers high degradation due to the
effects of channel impairments. Specifically, the channel
impairments, which include Inter-Symbol-Interference and the
leakage of I/Q modulated signals, which are sent over each of the
RF carriers, significantly degrades the Bit Error Rate (BER)
performance. The ISI is caused by the change of bandwidth of the
frequencies of specific symbols, spilling over to the next set of
symbols, or to the previous set of symbols. The leakage of I/Q
signals on each other is caused by the imperfect phase alignment
between the transmit and receiver carrier phases. In OFDM systems,
the leakage of I/Q signals can be more predominant in wireless
channels as opposed to wireline channels. In a small number of
carrier based systems, both wireline and wireless channels
experience leakage of I/Q due to imperfect phase imbalance. As the
transmit systems carry large data rates, the sensitivity to these
channel impairments become significant.
[0007] In accordance with the invention described herein, an
Artificial Neural Network (ANN) based demodulator is shown that
handles the ISI and I/Q leakage due to phase imbalance as a single
apparatus. The novel design of this demodulator simplifies the
adaptive demodulator complexity and improves the data recovery
process significantly in terms of Bit Error Rates (BER).
[0008] While the invention is applicable to broader transmission
channels, the preferred embodiment is a system that has a small
number of RF carriers for transmission over a wideband channel.
[0009] Traditional systems use two different systems to handle
these two channel impairments, where each system requires separate
training time during initialization. When both impairments are
handled separately with each requiring its own training time, the
computation time to optimize the design with appropriate correction
coefficients increases. Also, there could be bottlenecks in the
design process to achieve optimal system performance, when these
two impairments are handled sequentially, one after another. The
teaching of this invention is directed to an integrated demodulator
that avoids this pitfall by simultaneously handling both I/Q
imbalance and ISI with a single training sequence. This process
develops the necessary coefficients for an ANN demodulator to
achieve optimum performance.
[0010] In summary, this invention teaches the design of an
Artificial Neural Network based Demodulator that achieves the
following functions at the receiver:
[0011] 1. Compensates for the I/Q imbalance due to carrier phase
miss-alignment between the transmitter and receiver
[0012] 2. Equalizes the ISI introduced by the channel
[0013] 3. Equalizes the ISI introduced by the channel filter 4.
Recovers the original data which was used for modulation at the
sending side.
[0014] The proposed invention achieves significant advantages over
traditional methods of handling transmission impairments, for
example,
[0015] It reduces the computational complexity of the demodulation
process using single operation as opposed to multiple
operations.
[0016] It leads to more accurate and robust handling of channel
impairments at the receiver due to integrated operation instead of
sequential operations.
[0017] It increases the battery life of mobile apparatus
(particularly useful to handheld devices) by extending the mean
time before failure.
[0018] Equalization techniques broadly support handling
transmission impairments over different channels: wireline
communications or wireless communications or highly dispersive
channels. The transmission impairments can be different in
different channels.
[0019] In wireline channels, the channel equalization is designed
to handle ISI and reflections. The concept of equalization relates
to the loss compensation for the equalizer as a figure of merit,
which is used to derive the performance of the data recovery at the
receiver. Since the distance between the sending side and the
receiving side is fixed, the channel characteristics are known `a
priori` and it is possible to a use a Minimum Mean Squared Error
(MMSE) equalizer to minimize the effect of ISI. When the channel
transfer function is unknown, it is imperative to use an adaptive
MMSE equalizer.
[0020] There are implementations of equalizers used to handle ISI
based on Least Mean Squared Error (LMSE). This equalizer performs
well in minimizing the effect of ISI as long as the phase variation
on the channel is low. Although a LMSE equalizer works well in a
minimum phase channel, its performance is very limited in a channel
with spectral nulls. In such cases, the convergence of an LMS
equalizer is not guaranteed and ISI effects cannot be
minimized.
[0021] Another alternative to handle the ISI problem is the use of
a Decision Feedback Equalizer (DFE). While the DFE outperforms the
LMS, it is more complex than the LMS equalizer. Furthermore the DFE
suffers from an error propagation problem and therefore is only
used at very high SNR scenario. The MMSE, LMSE and DFE equalizers
can only minimize the effect of ISI on the performance, but cannot
handle the I/Q phase alignment problem. The present invention of an
integrated demodulator which both equalizes ISI and compensates
I/Q/imbalance outperforms a LMSE equalizer even for non-minimum
phase variation in the channel.
[0022] In wireless channels, the channel equalization is more
complex when handling ISI due to rapid changes in channel behavior
because of mobility and channel fading. The channel can be modeled
as a highly dispersive channel and will require a more complex
operation to reduce or eliminate the ISI effects. These channels
tend to be more time invariant, but are adaptive and therefore, the
channel equalizers tend to be adaptive to compensate and adjust for
the slow variations of the channel.
[0023] Some of the equalization methods used to handle wireless
channels include:
[0024] a. The method to nullify or mitigate the effect of channel
response by employing a training period to initialize the channel
equalizer that has a simple adaptive system. Some techniques in
this category include also a blind equalization technique without a
training period by employing different and possible-to-estimate
channel characteristics.
[0025] b. For OFDM channels, which use a narrow frequency band, the
channel equalization reduces the problem to handle flat fading or a
frequency non-selective system.
[0026] c. For handling channel fading, techniques such as multiple
transmission of the same information over independent channels and
waiting for the fading to recede before sending have been
exploited. The ultimate measure is the improvement of probability
of error in fading channels.
[0027] In summary, there are many teachings to design channel
equalizers based on neural networks to handle one selective
parameter at a time. As such, designing the equalizer to handle
multiple effects on the channel is more optimum and robust than
handling parameters one at a time which can cause delay in
processing to achieve optimization.
[0028] The proposed teaching in this invention is to demonstrate
designing the inventive demodulator to handle the effects of more
than one parameter simultaneously with a single training sequence
while achieving optimum performance for data recovery at the
receiver.
SUMMARY OF THE INVENTION
[0029] The invention described herein is directed to a neural
network based demodulator for use in a communication system,
wherein information sent over the communication channel can be
impaired by I/Q/imbalance and Inter Symbol Interference. The neural
network based demodulator functions to simultaneously compensate
for the I/Q imbalance and to equalize the Inter Symbol Interference
after a single integrated training step.
BRIEF DESCRIPTION OF THE DRAWINGS
[0030] FIG. 1 is a block schematic diagram of a high-speed data
transmission system.
[0031] FIG. 2 is an illustration of TXSRF using LO exponent of 4096
and RXSRF with LO exponent of 1.
[0032] FIG. 3 is a schematic of a neural network performing I/Q
balancing and equalization simultaneously.
DETAILED DESCRIPTION
[0033] The objective of the present invention is to provide a
design for an Artificial Neural Network (ANN) based Demodulator
that consists of the following elements which operate as an
integrated process for ISI handling and I/Q channel phase
imbalance: Sending side:
[0034] TXSRF (Pulse Shaper) Receiving side:
[0035] I/Q Demodulator
[0036] RXSRF (Pulse Matching)
[0037] ANN Demodulator with I/Q Balancer
[0038] ANN Equalizer
[0039] The preferred embodiment for the present invention is an ANN
Demodulator for use in a system such as described in U.S. Pat. Nos.
5,956,372 and 8,233,564, although the use of the present invention
is not limited to this preferred embodiment.
[0040] FIG. 1 shows the overall block schematic of the system for
high-speed data transmission that includes the components of the
ANN based Demodulator.
[0041] (1) TXSRF as a Pulse Shaper
[0042] The TXSRF which is part of the earlier patents (U.S. Pat.
Nos. 5,956,372 and 8,233,564) as well as this application's parent,
takes the Pulse Amplitude Modulated sinusoids and creates spike
voltages. In the referenced patents and patent application, the
spike voltages are created by having very high re-circulation for
each symbol in the order of a million samples per second. From a
practical implementation, the re-circulation rate is kept at 64
mega samples per second while the Local Oscillator has an exponent
of 4096. The TXSRF which creates the spikes is referred to as a
"Pulse Shaper" (see FIG. 1). FIG. 2 illustrates the modified TXSRF
with LO exponent of 4096. Exponent implies that the LO output is
raised to the power of 4096. Since the implementation of TXSRF can
be done using a look up table with all possible inputs and all
possible outputs, an LO exponent of this magnitude can be
implemented more realistically.
[0043] The output of TXSRF goes through a Transmit Band Pass Filter
(TXBPF) as shown in FIGS. 1 and 2.
[0044] (2) I/O Demodulator
[0045] Each RF carrier has two independent baseband data channels
which are orthogonal to each other (I and Q). They can be separated
and the demodulator extracts the baseband signals before being
processed by an individual RXSRF. The I and Q demodulated signals
are passed through RXBPF before processing by RXSRF.
[0046] (3) RXSRF (Pulse Matching)
[0047] The RXSRF is already shown in earlier U.S. Pat. Nos.
5,956,372 and 8,233,564, as well as this application's parent. It
is critical that the LO exponent is not increased in the RXSRF. The
input to the RXSRF is a combination of multiple signals from
different TXSRFs with added channel noise and when the exponent of
LO is raised, even though the signals spike, the channel noise can
exacerbate and cause degradation in the BER while recovering the
data. Therefore, the LO exponent of RXSRF is kept at 1 as shown in
FIG. 2.
[0048] (4) ANN Demodulator with I/Q Balancer
[0049] The performance of traditional digital modulation schemes
operating at high data rates suffers significant degradation due to
the effects of channel impairment. The important channel effects
include Inter Symbol Interference (ISI), and I/Q channel leakage
which is due to imperfect phase alignment between transmit and
receive carriers.
[0050] In the prior art, these two channel impairments are handled
separately, each requiring its own training time and computation
time to come up with appropriate correction coefficients.
[0051] For high data rate operation, this can create a bottleneck
resulting in poorer performance when each an impairment is handled
before the other. For example, when I/Q phase balancing is achieved
before handling ISI by using an equalizer, the equalizer may not
operate optimally with respect to performance errors. On the other
hand, if the equalizer is designed optimally, I/Q balancing may
fail resulting in higher performance errors.
[0052] The present invention teaches simultaneous handling of both
the impairments as part of the inventive ANN demodulator to achieve
optimal BER performance. The ANN demodulator handles the following
operations at the receiver:
[0053] 1. Compensates for the I/Q imbalance due to carrier phase
miss-alignment between the transmitter and receiver
[0054] 2. Equalizes the ISI introduced by the channel
[0055] 3. Equalizes the ISI introduced by the channel filter
[0056] 4. Recovers the original modulated symbols for recovering
the data.
[0057] The proposed invention uses the same training sequence for
handling the impairments simultaneously.
[0058] This invention simultaneously handles both I&Q balancing
and ISI with the same training sequence. The teachings of the
design of the neural network and the training algorithm for matched
filter application is presented in patent application U.S. Ser. No.
14/312,072, which is the parent of this application. The proposed
invention extends the teachings of U.S. Ser. No. 14/312,072 on the
ANN match filter described therein to a combined matched filtering,
equalization and I/Q balancing algorithm.
[0059] This invention reduces the training time required for the
digital signal processing of the algorithms for match filtering,
equalization and I/Q balancing. The training time taken for the
integrated process is significantly lower compared to processing
each of these algorithms independently. In addition, the
computational complexity of the demodulation process which is a
combination of multiple processes reduces to a single operation.
This reduction in complexity will increase the battery life of a
hand held device and extend the mean time before failure of the
device. In more general terms, this invention will increase the
versatility and agility of a digital communication receiver. Also
it leads to a more robust handling of channel impairments by the
receivers.
[0060] (5) ANN Demodulator
[0061] The demodulator is a combined I/Q balancer and equalizer
i.e. it has the capability of handling both interference and ISI
cancellation.
[0062] Assuming that a complex modulated data stream at the
transmitter is given by
x.sup.t=x.sub.I.sup.t+jx.sub.Q.sup.t (1)
[0063] After passing through a complex channel with channel matrix
given by
H=H.sub.I+jH.sub.Q (2)
[0064] The channel output Y is given as the product of X and H
yielding
y=Hx.sup.t+n=y.sub.I+jy.sub.Q+n=H.sub.Ix.sub.I.sup.t+jH.sub.Q.sup.t+n
(3)
[0065] Where n is the noise vector.
[0066] When the orthogonality of I and Q is lost due to imperfect
phase synchronization, the new received signal becomes a linear
combination of the I/Q components, i.e.
y'=y'.sub.I+jy'.sub.Q+n (4)
[0067] Where
y'.sub.I=ay.sub.I-by.sub.Q
y'.sub.Q=by.sub.I+ay.sub.Q (5)
[0068] With the parameters a=cos.theta. and b=sin.theta., and
.theta. is the phase angle difference between transmit carrier and
the reference LO carrier.
[0069] In traditional systems, the I/Q imbalance is first taken
care of before the channel equalization. This could be achieved in
two ways:
[0070] 1. Using carrier training: During the training period, an
unmodulated carrier is sent from the transmitter to the receiver.
Based on the received signal at the output of the matched filters,
.theta. could be determined using equation (5) and the receiver LO
phase can be adjusted accordingly to make sure that .theta. becomes
zero thereby isolating I from Q channels. This method is used when
the channel is (quasi) stationary.
[0071] 2. Using real time phase recovery: Alternatively, when the
channel varies more frequently, y.sub.I and y.sub.Q can be
extracted from y'.sub.I and y'.sub.Q without explicitly obtaining
the phase difference .theta.. This method comes with an additional
computational complexity during the data recovery stage.
[0072] After the I/Q channels are balanced, each of the recovered I
and Q data (y.sub.I and y.sub.Q) are then equalized independently
to recover the originally transmitted baseband modulated symbols
x.sub.I.sup.t and x.sub.Q.sup.t. Since H is not known a priori,
there's another training sequence required in order to estimate the
channel impulse response and then determine the equalizer
coefficient. The computed equalizer coefficients can then be used
to recover x.sub.I.sup.t and x.sub.Q.sup.t.
[0073] The overall computational complexity by the above method
introduces more real estate in the hardware as well as adds more
computational time.
[0074] In contrast, in the proposed teaching, the inventive
demodulator described herein uses only a single training period,
where the equalizer function iteratively computes efficient
coefficients that can be used to balance I/Q channels and perform
equalization at the same time. During the training stage, both I/Q
data are fed into the neural network as inputs (shown in FIG. 3).
The neural network shown in FIG. 3 is the same two layer
configuration shown in FIG. 5 of parent application Ser. No.
14/312,072. The neural network will then determine the coefficients
to create an appropriate model that performs I/Q balancing and
equalization at the same time. Since the computational complexity
of training a single neural network is similar to that of training
a traditional equalizer, the second stage training process has now
been reduced into a single one. Also, only one neural network
equalizer is required for both I/Q channels.
[0075] As set forth in parent application Ser. No. 14/312,072, the
neural network training process determines the coefficients w1, b1,
w2, b2 such that the mean squared error between the transmitted
symbols x.sup.t and the equalized symbols x.sup.r (where
x.sup.r=x.sub.I.sup.r+jx.sub.Q.sup.r is the output of the neural
network demodulator is minimized. That is, the optimization problem
can be defined as:
argmin w 1 , b 1 , w 2 , b 2 E [ ( x r - x t ) 2 ] ( 5 )
##EQU00001##
The neural network output is calculated from the following
mathematical steps: The neural network input is the vector
concatenation of the received I & Q data, i.e.
P=[y'.sub.t; y'.sub.Q] (6)
N=w1.times.P+b1 (7)
A1=TF(N)=max(-1, min(1, N)) (8)
A2=w2.times.A1+b2 (9)
x.sup.r=A2[x.sub.I.sup.r]+j A2[x.sub.Q.sup.r] (10)
[0076] Where A2[z] is the subset of A2 that correspond to the
z.
[0077] It is worth mentioning that the neural network input signal
could be sampled at the symbol rate and therefore the neural
network equalizer will function similar to a symbol-by-symbol
equalizer or the input could be oversampled giving rise to a
fractionally spaced equalizer. In the case where input sampling
rate is the same as the sampling rate at the receiver input, the
neural network demodulator can flexibly add matched filtering to
its functions. In fact it is the ability of the neural network to
combine matched filtering, I/Q balancing and equalization that lead
to the name "neural network demodulator".
[0078] The following section compares the neural network equalizer
function to the traditional transversal (LSM/RLS) equalizer:
[0079] The neural network equalizer can be seen as a two stage
transversal equalizer. Therefore the residual ISI from the first
layer is canceled in the second layer which makes it more superior
to traditional transversal equalizers.
[0080] The neural network equalizer can be trained with varying
length of input vector to speed up its convergence depending on how
fast the channel is varying whereas as traditional transversal
equalizer can only train one input sample at a time. In fact, due
to the stochastic nature of the input vector, traditional
transversal equalizers may never converge or be able to capture the
underlying ISI signature of the channel.
[0081] Also, the neural network equalizer contains a transfer
function TF(N) between the two layers to ensure that the training
converges and does not diverge or get stuck in a sub-optimal local
minimal
[0082] The proposed invention uniquely exploits the flexibility
that the neural network offers to handle multiple tasks faster
which has never been exploited in traditional equalizers as
designed for use in digital communication systems.
[0083] Although the present invention has been described in
conjunction with specific embodiments, those skilled in the art of
the present invention will appreciate that modifications and
variations can be made without departing from the scope and the
spirit of this invention. Such modifications and variations are
envisioned to be within the scope of the amended claims.
* * * * *