 |
Description  |
|
|
BACKGROUND AND SUMMARY
The invention relates to active acoustic attenuation systems, and more
particularly to those systems providing sound cancellation in the presence
of feedback sound from a compensating speaker or transducer, which sound
is coupled back into the input and hence into the cancelling loop.
Prior feedback cancellation systems use a filter to compensate for feedback
sound from the speaker to the input microphone. It is desirable that this
filter be adaptive in order to match the changing characteristics of the
feedback path. Prior systems will successfully adapt only for broad band
noise input signals because the system input is uncorrelated with the
output of the feedback cancellation filter. Uncorrelated signals average
to zero over time. However, if the input noise contains narrow band noise
such as a tone having a regular periodic or recurring component, as at a
given frequency, the filter output will be correlated with the system
input and will not converge. The filter may thus be used adaptively only
in systems having exclusively broad band input noise.
Most practical systems, however, do experience narrow band noise such as
tones in the input noise. The noted filter cannot be adaptively used in
such systems. To overcome this problem, and as is known in the prior art,
the filter has been pre-trained off-line with broad band noise only. This
pre-adapted filter is then fixed and inserted into the system as a fixed
element which does not change or adapt thereafter.
A significant drawback of the noted fixed filter is that it cannot change
to meet changing feedback path characteristics, such as temperature or
flow changes in the feedback path, which in turn change the speed of
sound. During the pre-training process, the filter models a pre-determined
set of given parameters associated with the feedback path, such as length,
etc. Once the parameters are chosen, and the filter is pre-adapted, the
filter is then inserted in the system and does not change thereafter
during operation. This type of fixed filter would be acceptable in those
systems where feedback path characteristics do not change over time.
However, in practical systems the feedback path does change over time,
including temperature, flow, etc.
It is not practical to always be shutting down the system and re-training
the filter every time the feedback path conditions change, nor may it even
be feasible where such changes occur rapidly, i.e., by the time the system
is shut down and the filter re-trained off-line, the changed feedbackk
path characteristic such as temperature may have changed again. For this
reason, the above-noted fixed filter is not acceptable in most practical
systems.
There is thus a need for truly adaptive feedback cancellation in a
practical active acoustic attenuation system, where the characteristics of
the feedback path may change with time. A system is needed wherein the
feedback is adaptively cancelled on-line for both broad band and narrow
band noise without dedicated off-line pre-training, and wherein the
cancellation further adapts on-line for changing feedback path
characteristics such as temperature and so on.
BRIEF DESCRIPTION OF THE DRAWINGS
Prior Art
FIG. 1 is a schematic illustration of an active acoustic attenuation system
known in the prior art.
FIG. 2 is a block diagram of the embodiment in FIG. 1.
FIG. 3 is a schematic illustration of a feedback cancellation active
acoustic attenuation system known in the prior art.
FIG. 4 is a block diagram of the embodiment in FIG. 3.
Present Invention
FIG. 5 is a schematic illustration of acoustic system modeling in
accordance with the invention.
FIG. 6 is a block diagram of the system in FIG. 5.
FIG. 7 is one embodiment of the system in FIG. 6.
FIG. 8 is another embodiment of the system in FIG. 6.
FIG. 9 is a further embodiment of the system in FIG. 6.
FIG. 10 is a schematic illustration of the system in FIG. 7.
FIG. 11 is a schematic illustration of the system in FIG. 9.
DETAILED DESCRIPTION
Prior Art
FIG. 1 shows a known prior art acoustic system 2 including a propagation
path or environment such as a duct or plant 4 having an input 6 for
receiving input noise and an output 8 for radiating our outputting output
noise. The input noise is sensed with an input microphone 10 and an input
signal is sent to controller 9 which drives unidirectional speaker array
13 which in turn injects cancelling sound into duct or plant 4 which sound
is optimally equal in amplitude and opposite in sign to the input noise to
thus cancel same. The combined noise is sensed with an output microphone
16 whichprovides an error signal fed to controller 9 which then outputs a
correction signal to speaker array 13 to adjust the cancelling sound. The
error signal at 15 is typically multiplied with the input signal at 11 by
multiplier 17 and the result provided as weight update signal 19, for
example as discussed in Gritton and Lin "Echo Cancellation Algorithms",
IEEE ASSP Magazine, April 1984, pp. 30-38. In some prior art references,
multiplier 17 is explictly shown, and in others the multiplier 17 or other
combination of signals 11 and 15 is inherent or implied in controller 9
and hence multiplier or combiner 17 may be deleted in various references,
and such is noted for clarity. For example, FIG. 2 shows the deletion of
such multiplier or combiner 17, and such function, if necessary, may be
implied in controller 9, as is understood in the art.
Speaker array 13 is unidirectional and emits sound only to the right in
FIG. 1, and does not emit sound leftwardly back to microphone 10, thus
preventing feedback noise. The particular type of unidirectional speaker
array shown is a Swinbanks type having a pair of speakers 13a and 13b
separated by a distance L. The input to speaker 13b is an inverted version
of the input to speaker 13a that has been delayed by a time .tau.=L/c
where c is the speed of sound. This arrangement elminates acoustic
feedback to microphone 10 over a limited frequency range. The time delay
.tau. must be adjusted to account for changes in sound speed due to
temperature variations. Other types of unidirectional speakers and arrays
are also used, for example as shown in "Historical Review and Recent
Development of Active Attenuators", H. G. Leventhall, Acoustical Society
of America, 104th Meeting, Orlando, November, 1982, FIG. 8. In another
system, a unidirectional microphone or an array of microphones is used at
10, to ignore feedback noise. Other methods for eliminating the feedback
problem are also used, such as a tachometer sensing rotational speed, if a
rotary source provides the input noise, and then introducing cancelling
sound according to sensed RPM, without the use of a microphone sensing
input noise at 10. Other systems employ electrical analog feedback to
cancel feedback sound. Others employ a fixed delay to cancel known delayed
feedback sound.
Acoustic system 4 is modeled by controller model 9 having a model input
from input microphone 10 and an error input from output microphone 16, and
outputting a correction signal to speaker array 13 to introduce cancelling
sound such that the error signal approaches a given value, such as zero.
FIG. 2 shows the modeling, with acoustic system 4 shown at the duct or
plant P, the modeling controller 9 shown at P', and the summation thereof
shown at 18 at the output of speaker array 13 where the sound waves mix.
The output of P is supplied to the plus input of summer 18, and the output
of P' is supplied to the minus input of summer 18. Model 9, which may use
the least means square (LMS) algorithm, adaptively cancels undesirable
noise, as is known, and for which further reference may be had to "Active
Adaptive Sound Control in a Duct: A Computer Simulation", J. C. Burgess,
Journal of Acoustic Society of America, 70(3), September, 1981, pp.
715-726, to Warnaka et al U.S. Pat. No. 4,473,906, and to Widrow, Adaptive
Filters, "Aspects of Network and system Theory", edited by R. E. Kalman
and N. DeClaris, Holt, Reinhart and Winston, New York, 1971, pp. 563-587.
The system of FIGS. 1 and 2 operates properly when there is no feedback
noise from speaker array 13 to input microphone 10.
It is also known to provide an omnidirectional speaker 14, FIG. 3, for
introducing the cancelling sound, and to provide means for compensating
feedback therefrom to the input microphone. As seen in FIG. 3, the
cancelling sound introduced from omnidirectional speaker 14 not only mixes
with the output noise to cancel same, but also travels leftwardly and is
sensed at input microphone 10 along feedback path 20, as shown in FIG. 3
where like reference numerals are used from FIG. 1 where appropriate to
facilitate clarity. In one known system for cancelling feedback, as shown
in Davidson Jr. et al U.S. Pat. No. 4,025,724, the length of the feedback
path is measured and then a filter is set accordingly to have a fixed
delay for cancelling such delayed feedback noise. In another known sysem
for cancelling feedback, a dedicated feedback control 21 in the form of a
filter is provided, for example as shown in "Active Noise Reduction
Systems in Ducts", Tichy et al, ASME Journal, November, 1984, page 4, FIG.
7, and labeled "adaptive uncoupling filter". Feedback control filter 21 is
also shown in the above noted Warnaka et al U.S. Pat. No. 4,473,906 as
"adaptive uncoupling filter 75" in FIGS. 14 and 15, and in "The
Implentation of Digital Filters Using a Modified Widrow-Hoff Algorithm For
the Adaptive Cancellation of Acoustic Noise", Poole et al, 1984 IEEE, CH
1945-5/84/0000-0233, pp. 21.7.1-21.7.4. Feedback control filter 21
typically has an error signal at 26 multiplied with the input signal at 24
by multiplier 27 and the result provided as weight update signal 29.
Feedback control or adaptive uncoupling filter 21 is pre-trained off-line
with a dedicated set of parameters associated with the feedback path. The
filter is pretrained with broad band noise before the system is up and
running, and such predetermine dedicated fixed filter is then inserted
into the system.
In operation in FIG. 3, controller 9 is a least mean square (LMS) adaptive
filter which senses the input from microphone 10 and outputs a correction
signal to speaker 14 in an attempt to drive the error signal from
microphone 16 to zero, i.e., controller 9 continually adaptively changes
the output correction signal to speaker 14 until its error input signal
from microphone 16 is minimized. Feedback control filter 21 has an input
at 24 from the output of controller 9.
During off-line pre-training, switch 25 is used to provide filter 21 with
an error input at 26 from summer 28. During the off-line pre-training,
switch 25 is in its upward position to contact terminal 25a. During this
pre-training, broad band noise is input at 35, and feedback control 21
changes its output 30 until its error input at 26 is minimized. The output
30 is summed at 28 with the input from microphone 10, and the result is
fed to controller 21. Feedback control 21 is pre-trained off-line to model
feedback path 20, and to introduce a cancelling component therefor at 30
to summer 28 to remove such feedback component from the input to
controller 9 at 32. LMS adaptive filter 21 is typically a transversal
filter and once its weighting coefficients are determined during the
pre-training process, such coefficients are kept fixed thereafter when the
system is up and running in normal operation.
After the pre-training process, switch 25 is used to provide an input to
controller 9, and the weighting coefficients are kept constant. After the
pre-training process and during normal operation, switch 25 is in its
downward position to contact terminal 25b. The system is then ready for
operation, for receiving input noise at 6. During operation, feedback
control 21 receives no error signal at 26 and is no longer adaptive, but
instead is a fixed filter which cancels feedback noise in a fixed manner.
The system continues to work even if narrow band noise such as a tone is
received at input 6. However, there is no adaptation of the filter 21 to
changes in the feedback path due to temperature variations and so on.
FIG. 4 shows the system of FIG. 3 with feedback path 20 summed at 34 with
the input noise adjacent microphone 10. Fixed feedback control
cancellation filter 21 is shown at F', and adaptive controller 9 at P'.
Adaptive controller 9 at P' models the duct or plant 4 and senses the
input at 32 and outputs a correction signal at 35 and varies such
correction signal until the error signal at 36 from summer 18 approaches
zero, i.e., until the combined noise at microphone 16 is minimized. Fixed
filter 21 at F' models the feedback path 20 and removes or uncouples the
feedback component at summer 28 from the input 32 to filter 9. This
prevents the feedback component from speaker 14 from being coupled back
into the input of the system model P'. As above noted, the error signal at
26 is only used during the training process prior to actual system
operation.
It is also known that propagation delay between speaker 14 and microphone
16 if any, may be compensated by incorporating a delay element in input
line 33 to compensate for the inherently delayed error signal on line 36.
Feedback model F' at filter 21 will successfully adapt for broad band noise
because the system input is uncorrelated with the output of the feedback
cancellation filter. Filter 21 may thus model the predetermined feedback
path according to the preset feedback path characteristic. However, if the
input noise contains any narrow band noise such as a tone having a regular
periodic or recurring component, as at a given frequency, the output of
filter 21 will be correlated with the system input and will continue to
adapt and not converge. Filter 21 may thus be used adaptively only in
systems having exclusively broad band input noise. Such filter is not
amenable to systems where the input noise may include any narrow band
noise.
Most practical systems do have narrow band noise in the input noise. Thus,
in practice, filter 21 is pre-adapted and fixed to a given set of
predetermined feedback path characteristics, and does not change or adapt
to differing feedback path conditions over time, such as temperature, flow
rate, and the like, which affect sound velocity. It is not practical to
always be retraining the filter every time the feedback path conditions
change, nor may it even be feasible where such changes occur rapidly,
i.e., by the time the system is shut down and the filter retrained
off-line, the changed feedback path characteristic such as temperature may
have changed again.
Thus, the feedback control system of FIGS. 3 and 4 is not adaptive during
normal operation of the system. Filter 21 must be pre-trained off-line
with broad band noise and then fixed, or can only be used adaptively
on-line with exclusively broad band noise input. These conditions are not
practical.
There is a need for truly adaptive feedback cancellation in an active
attenuation system, wherein the feedback is adaptively cancelled on-line
for both broad band and narrow band noise without dedicated off-line
pre-training, and wherein the cancellation further adapts on-line for
changing feedback path characteristics such as temperature and the like.
Present Invention
FIG. 5 shows a modeling system in accordance with the invention, and like
reference numerals are used from FIGS. 1-4 where appropriate to facilitate
clarity. Acoustic system 4, such as a duct or plant, is modeled with an
adaptive filter model 40 having a model input 42 from input microphone or
transducer 10 and an error input 44 from output microphone or transducer
16, and outputting a correction signal at 46 to omnidirectional speaker or
transducer 14 to introduce cancelling sound or acoustic waves such that
the error signal at 44 approaches a given value such as zero. In FIG. 5,
sound from speaker 14 is permitted to travel back along feedback path 20
to input microphone 10 comparably to FIG. 3, and unlike FIG. 1 where such
feedback propagation is prevented by unidirectional speaker array 13. The
use of an omnidirectional speaker is desirable because of its availability
and simplicity, and because it eliminates the need to fabricate a system
of speakers or other components approximating a unidirectional
arrangement.
In accordance with the invention, feedback path 20 from transducer 14 to
input microphone 10 is modeled with the same model 40 such that model 40
adaptively models both acoustic system 4 and feedback path 20. The
invention does not use separate on-line modeling of acoustic system 4 and
off-line modeling of feedback path 20. In particular, off-line modeling of
the feedback path 20 using broad band noise to pretrain a separate
dedicated feedback filter is not necessary. Thus, in the prior art of FIG.
4, the feedback path F at 20 is modeled separately from the direct path 4
at plant P, with a separate model 21 at F' pretrained solely to the
feedback path and dedicated thereto as above noted. In the present
invention, the feedback path is part of the model 40 used for adaptively
modeling the system.
FIG. 6 shows the system of FIG. 5 in accordance with the invention, wherein
acoustic system 4 and feedback path 20 are modeled with a single filter
model 40 having a transfer function with poles used to model feedback path
20. This is a significant advance over the art because it recognizes that
individual finite impulse response (FIR) filters shown in FIGS. 3 and 4
are not adequate to truly adaptively cancel direct and feedback noise.
Instead, a single infinite impulse respone (IIR) filter is needed to
provide truly adaptive cancellation of the direct noise and acoustic
feedback. In accordance with the invention, the acoustic system and the
feedback path are modeled on-line with an adaptive recursive filter model.
Since the model is recursive, it provides the IIR characteristic present
in the acoustic feedback loop wherein an impulse will continually feed
upon itself in feedback manner to provide an infinite response.
As noted in the above referenced Warnaka et al U.S. Pat. No. 4,473,906,
column 16, lines 8+, the adaptive cancelling filter in prior systems is
implemented by a transversal filter which is a non-recursive finite
impulse response filter. These types of filters are often referred to as
all-zero filters since they employ transfer functions whose only roots are
zeros, "VLSI Systems Designed for Digital Signal Processing", Bowen and
Brown, Vol. 1, Prentice Hall, Englewood Cliffs, N.J., 1982, pp. 80-87. To
adaptively model acoustic system 4 and feedback path 20 with a single
filter model 40 requires a filter with a transfer function containing both
zeros and poles. Such poles and zeros are provided by a recursive IIR
algorithm. The present invention involves providing an IIR recursive
filter model to adaptively model acoustic system 4 and feedback path 20.
This problem has been discussed by Elliot and Nelson in I.S.V.R. Technical
Report No. 127, Southampton University, England, published in U.S.
Department of Commerce, National Technical Information Service, Bulletin
No. PB85189777, April 1984. In discussing the use of recursive models for
use in active attenuation systems, Elliot et al note, page 37, that the
number of coefficients used to implement the direct and feedback modeling
can desirably be kept to a minimum, however they further note that there
is "no obvious method" to use in obtaining the responses of the recursive
structure. In the conclusion on page 54, last paragraph, Elliott et al
note that "no procedure has yet been developed for adapting the
coefficients of a recursive IIR filter to obtain the best attenuation".
The present invention provides a system that solves this problem and
adaptively determines these coefficients in a practical system that is
effective on broad band as well as narrow band noise.
The poles of the transfer function of the model 40 result in a recursive
characteristic that is necessary to simultaneously model the acoustic
system 4 and the feedback path 20. The response of model 40 will feedback
upon itself and can be used to adaptively cancel the response of the
feedback path 20 which will also feedback upon itself. In contrast, in an
FIR filter, there is no feedback loop but only a direct path through the
system and only zeros are possible, as in the above noted Tichy et al
article and Warnaka et al patent, i.e., the zeros of the numerator of the
transfer function. Thus, two individual models must be used to model the
acoustic system 4 and feedback path 20.
For example, in Tichy et al and Warnaka et al, two independent models are
used. The feedback path is modeled ahead of time by pre-training the
feedback filter model off-line. In contrast, in the present invention, the
single model adapts for feedback on-line while the system is running,
without pre-training. This is significant because it is often impossible
or not economically feasible to retrain for feedback every time the
feedback path characteristics change, e.g., with changing temperature,
flow rate, etc. This is further significant because it is not known when
narrow band noise such as a tone may be included in the input noise, and
must be adaptively accommodated and compensated for.
FIG. 7 shows one form of the system of FIG. 6. The feedback element B at 22
is adapted by using the error signal at 44 as one input to model 40, and
the correction signal at 46 as another input to model 40, together with
the input at 42. The direct element A at 12 has an output summed at 48
with the output of the feedback element B at 22 to yield the correction
signal at 46 to speaker or transducer 14 and hence summer 18.
In FIG. 8, the input to feedback element B at 22 is provided by the output
noise at 50 instead of the correction signal at 46. This is theoretically
desirable since the correction signal at 46 tends to become equal to the
output noise at 50 as the model adapts. Improved performance is thus
possible through the use of the output noise 50 as the input to the
feedback element B from the beginning of operation. However, it is
difficult to measure the output noise without the interaction of the
cancelling sound from speaker 14. FIG. 9 shows a particularly desirable
implementation in accordance with the invention enabling the desired
modeling without the noted measurement problem. In FIG. 8, the feedback
element is adapted at B using the error signal at 44 from the output
microphone as one input to model 40, and the output noise at 50 as another
input to model 40. In FIG. 9, the error signal at 44 is summed at summer
52 with the correction signal at 46, and the result is provided as another
input at 54 to model 40. This input 54 is equal to the input 50 shown in
FIG. 8, however it has been obtained without the impractical acoustical
measurement required in FIG. 8. In FIGS. 7-9, one of the inputs to model
40 and to feedback element B component 22 is supplied by the overall
system output error signal at 44 from output microphone 16. The error
signal at 44 is supplied to feedback element B through multiplier 45 and
multiplied with input 51, yielding weight update 47. Input 51 is provided
by correction signal 46, FIG. 7, or by noise 50, FIG. 8, or by sum 54,
FIG. 9. The error signal at 44 is supplied to direct element A through
multiplier 55 and multiplied with input 53 from 42, yielding weight update
49.
The invention enables in its preferred embodiment the use of a recursive
least mean square (RLMS) algorithm filter, for example "Comments on `An
Adaptive Recursive LMS Filter`", Widrow et al, Proceedings of the IEEE,
Vol. 65, No. 9, September 1977, pp. 1402-1404, FIG. 2. The invention is
particularly desirable in that it enables the use of this known recursive
LMS algorithm Filter. As shown in FIG. 10, illustrating the system of FIG.
7, the direct element A at 12 may be modeled by an LMS filter, and the
feedback element B at 22 may be modeled with an LMS filter. The adaptive
recursive filter model 40 shown in the embodiment of FIG. 10 is known as
the recursive least mean square (RLMS) algorithm.
In FIG. 11, showing the system in FIG. 9, the feedback path 20 is modeled
using the error signal at 44 as one input to model 40, and summing the
error signal at 44 with the correction signal at 46, at summer 52, and
using the result at 54 as another input to model 40.
The delay, if any, in output 8 between speaker 14 and microphone 16, may be
compensated for by a comparable delay at the input 51 to LMS filter 22
and/or at the input 53 to LMS filter 12.
The present invention thus models the acoustic system and the feedback path
with an adaptive filter model having a transfer function with poles used
to model the feedback path. It is of course within the scope of the
invention to use the poles to model other elements of the acoustic system
in combination with modeling the feedback path. It is also within the
scope of the invention to model the feedback path using other
characteristics, such as zeros, in combination with the poles.
It is recognized that various equivalents, alternatives and modifications
are possible within the scope of the appended claims.
* * * * *
|
|
 |
|
 |
Description |
|
