JPH07160508A - Filter coefficient deciding method for adaptive filter - Google Patents

Abstract

PURPOSE:To accelerate the study of a filter coefficient by using a fuzzy theory concerning the filter coefficient deciding method for adaptive filter using a Filtered-X/NLMS algorithm based on the fuzzy theory. CONSTITUTION:A device temperature Ta, error signal E to be inputted to the coefficient arithmetic part of the adaptive filter, step gain K and impulse response sample value a(t) of an error diffusion system are respectively defined by a membership function, control rules with the device temperature Ta and the error signal E as an antecedent part and with the step gain K and the impulse response sample value a(t) as a consequent part are defined, the step gain K and the impulse response sample value a(t) are calculated from the states of the device temperature Ta and the error signal E by the fuzzy theory and based on that calculated result, the filter coefficient of the adaptive filter is decided.

Description

Detailed Description of the Invention

[0001]

FIELD OF THE INVENTION The present invention is based on the fuzzy theory of Fi.
The present invention relates to a method for determining a filter coefficient of an adaptive filter using the ltered-X / NLMS algorithm.

The filter coefficient determination method of the present invention can be used for any electronic equipment using an adaptive FIR digital filter (hereinafter, simply referred to as an adaptive filter) which uses a Filtered-X / NLMS algorithm as an adaptive algorithm of the filter coefficient. You can

[0003]

2. Description of the Related Art In recent years, in order to improve the working environment in an office or the like, it has been required to reduce the noise generated by information processing equipment and the like, and active noise control technology has attracted attention as this noise reduction technology. . The principle of this active noise control is to generate a pseudo noise with the same amplitude and opposite phase as the noise with an adaptive filter, and to cancel the noise by superimposing it on the noise. It can be expected to have a great effect on suppressing low-frequency noise that is difficult to suppress.

The main technique of this active noise control is an adaptive algorithm for calculating the filter coefficient of an adaptive filter for synthesizing pseudo noise having the same amplitude and opposite phase as noise, and the most typical adaptive algorithm is Filtered-. X /
The LMS (Least mean squares) method is known. LMS
In the method, the filter coefficient of the adaptive filter is calculated while being corrected so that the input of the error detection microphone placed at the position where noise reduction is sought, that is, the canceling noise (error) between noise and pseudo noise is minimized. To be done.

However, in active noise control,
There is a problem that the error can only be observed indirectly through the error detection microphone, and the transfer system that reaches the adaptive algorithm via this error detection microphone (this transfer system scatters the error in the time axis direction). Therefore, when the error scattering system here cannot be approximated to 1, it is necessary to pre-process the LMS method with a filter simulating this transmission system (referred to as a scattering estimation filter). This adaptive algorithm is Filtered-X / LMS
Is the law.

FIG. 6 shows a control system of an active noise control system using the Filtered-X / LMS method as an adaptive algorithm. As shown in FIG. 6, the noise elimination filter 1, the wraparound filter 2, the scattering estimation filter 3, the calculation control unit 4, the speaker 5, the noise detection microphone 6, the error detection microphone 7, and the like are configured. Reference numeral 8 denotes a noise erasing duct that serves as a path for erasing noise, and also serves as a duct for exhausting heat generated inside the device that is the noise source.

The noise detection microphone 6 is installed on the noise source side in the duct 8 to detect noise. The noise elimination filter 1 uses the noise detected by the microphone 6 as a reference signal to generate a pseudo noise having the same amplitude and opposite phase as the noise.
It is a filter and simulates a noise transmission system A from the microphone 6 to the speaker 5.

The speaker 5 is for electrically / acoustic converting the pseudo noise generated by the noise elimination filter 1, and superimposing the pseudo noise that has become acoustic on the noise that has propagated through the transmission system A in the duct 8. The error detection microphone 7 is a microphone that detects an error sound (= noise-pseudo noise) that cannot be completely erased in order to update the filter coefficient of the noise elimination filter 1. The error signal from the microphone 7 is input to the arithmetic control unit 4.

In the adaptive algorithm of the filter coefficient of the noise canceling filter 1, the scattering estimation filter 3 passes through the error detection microphone 7 and reaches the calculation control unit 4 in the transmission system B.
This is an FIR filter for simulating this transfer system B in order to consider the influence of the above-mentioned error scattering system.

Based on the output of the scattering estimation filter 3 and the error signal of the microphone 7, the arithmetic control unit 4 executes Filtered-X.
A circuit for performing and controlling calculation for updating the filter coefficient of the noise elimination filter 1 according to the / LMS method.

The sneak-in prevention filter 2 is a filter for simulating a transmission system C that reversely traces the noise transmission system A. In the noise reduction device, the pseudo noise emitted from the speaker 5 propagates through the transmission system C and enters the noise detection microphone 6, and when it is input to the noise elimination filter 1, the noise elimination filter 1 accurately eliminates the noise. Since the generation of sound is hindered, the transfer system C is simulated by the sneak filter 2, the pseudo noise from the noise elimination filter 1 is generated through the sneak filter 2, and the pseudo sneak sound is generated by the noise detection microphone 6. Detection signal (= noise + wraparound sound)
The effect of the rounding sound passing through the transmission system C is removed by subtracting from.

As a method that further advances this Filtered-X / LMS method, Filtered-X / NLMS (Normalized LMS) is used.
MS) method. Concerning the convergence condition of this Filtered-X / NLMS method, the Institute of Electronics, Information and Communication Engineers, IEEJ, EA9
3-10, May 1993, pages 33-38, Fujii et al., "Study on convergence of Filtered-X / NLMS method".

According to the report, Filtered-X / N shown in FIG.
As shown in the estimation system of the adaptive filter according to the LMS method, Filtered-X / NLMS method, the unknown system output g _i and the output G _i (= pseudo noise noise cancellation filter 1 of the adaptive filter (= transmission system A) ) With the error E
_i (= error signal of the error detection microphone 7) is one of the filter coefficient successive update algorithms applied to a system that can be obtained only through the error scattering system (= transmission system B). The point is that a scattering estimation filter simulating an error scattering system for diffusing E _j in the direction of the time axis is placed before the NLMS method.

[0014] update equation of the filter coefficients of the adaptive filter in this system, expression yielding an estimate H _{j + 1} (m) at time j + 1 for the m-th tap coefficient of the adaptive filter, H j _{+ 1} (m) = H _j (m) + Ke _j Y _j (m) / ΣY _j ² (i) (1) However, Y _j is the output of the scattering estimation filter, e _j is the error E _j affected by the error scattering system,
K is a step gain, Σ is cumulative addition of i = 1 to I, and I is the number of taps of the adaptive filter.

This Filtered-X / NL is shown in FIG.
A T-th order cyclic filter expression of the MS method is shown, and a first-order cyclic filter expression of the NLMS method is shown in FIG. In the normal NLMS method, the first-order cyclic filter shown in FIG. Therefore, the estimation error is calculated as the sum of I estimation errors generated in the output of the first-order cyclic filter. Further, this sum is not an addition of the estimation error generated for each filter when considering an average virtual filter common to the individual first-order recursive filters, but an estimation error generated at the output of the I virtual filters. It can also be I times.

The concept of this virtual filter is Filtered-X
When the average filter coefficient α _j (t) of the virtual virtual T-th order cyclic filter common to m = 1 to I is calculated by applying it to the T-th order cyclic filter formed by the / NLMS method, α _j (t) = K [a ² (t) ΣX _j-1 ² (i) + Σa (t) X _j-1 (i) Σ _ut a (u) X _ju (i)] / [IΣY _j ² (i)] -It is expressed as (2). Here, a (t) is a sample value of the impulse response of the error scattering system, and Σ _ut is a cumulative addition of u = 0 to T−1 excluding u = t. That is, in the normal NLMS method, the coefficient of the virtual filter becomes a constant, whereas in the Filtered-
In the X / NLMS method, it will be given as a random variable. The filter coefficient of the noise elimination filter 1 can be determined based on the virtual filter coefficient α.

[0017]

According to the report, Filt
The convergence condition regarding the learning of the ered-X / NLMS method is that a stable low-pass filter with a gain of 1 is formed in the FIR filter, and the gain of the cyclic circuit with respect to the estimation error is less than 1. In order to satisfy this condition, the step gain K and the sampled value a (t) of the impulse response are held as fixed values as shown in the equation (2) of the filter coefficient α (t) of the virtual filter. Therefore, if there is a change in the environment such as temperature fluctuations or active maintenance, it is difficult to satisfy the convergence conditions with the above, and the convergence accuracy and convergence speed become undefined.
There is a problem that learning filter coefficients takes a very long time.

The present invention has been made in view of the above problems, and an object of the present invention is to speed up learning of filter coefficients in the Filtered-X / NLMS method by using fuzzy theory.

[0019]

FIG. 1 is a diagram illustrating the principle of the present invention. A filter coefficient determination method for an adaptive filter according to the present invention is a filter coefficient determination method for an adaptive filter that uses a Filtered-X / NLMS method considering an error scattering system as an algorithm for calculating a filter coefficient,
The device temperature Ta and the error signal E input to the coefficient calculation unit of the adaptive filter are defined by membership functions, and the step gain K and the impulse response sample value a of the error scattering system are defined.
(t) is defined by a membership function, and a control rule is defined in which the device temperature Ta and the error signal E are the antecedent part, and the step gain K and the impulse response sample value a (t) are the antecedent part. The step gain K and the impulse response sample value a (t) are obtained from the state of Ta and the error signal E by fuzzy inference, and the filter coefficient of the adaptive filter is determined based on the obtained result.

The filter coefficient determination method of the adaptive filter described above can be applied to an active noise control system in which pseudo noise having the same amplitude and opposite phase as noise is generated by the adaptive filter and the noise is suppressed by using the pseudo noise. , The error between the noise and the pseudo noise is used as the error signal.

[0021]

The filter coefficient of the adaptive filter is ultimately the step gain K and the impulse response sample value a of the error scattering system.
It can be decided based on (t). Therefore, the step gain and the impulse response sample value a (t) are obtained by fuzzy reasoning according to a predetermined control rule with the device temperature Ta and the error signal E as inputs. Thereby, the filter coefficient can be quickly determined.

[0022]

Embodiments of the present invention will be described below with reference to the drawings. Here, as an embodiment of the present invention, the filter coefficient determination method of the adaptive filter is applied to the active noise control system shown in FIG. Also Filtered-X
The estimation system of the filter coefficient by the / NLMS method is as shown in FIG. In the active noise control system, referring to the above equation (2) for determining the virtual filter coefficient α (t), this virtual filter coefficient α (t) is eventually calculated by the step gain K and the impulse response sample value a (t). It can be seen that the function, that is, α (t) = F [K, a (t)]. Therefore, if the step gain K and the impulse response sample value a (t) are obtained, the filter coefficient of the noise elimination filter 1 can be determined. Therefore, in the present invention, the step controller K and the impulse response sample value a are calculated by using the fuzzy theory in the arithmetic control unit 4.
Find the optimal value of (t).

FIG. 2 shows the concept of the logic for obtaining the step gain K and the impulse response sample value a (t) in the arithmetic control unit 4. The step gain K and the impulse response sample value a (t) can be considered as a function of the device temperature Ta and the error signal E from the error detection microphone 7. Therefore, the apparatus temperature Ta is fuzzy expressed by a membership function as shown in FIG. 3A, and three membership functions of low temperature, normal temperature and high temperature are defined as the state of the apparatus temperature Ta. Similarly, the error signal E is fuzzy expressed by a membership function as shown in FIG.
We define three membership functions with small error, ordinary error, and large error as states of.

Further, as shown in FIG. 4A, the step gain K is changed from a random variable to a fuzzy variable, and the step gain K is divided into three values: "small", "normal", and "large". Defined by the membership function of the state. Similarly, as shown in (B) of FIG. 4, the impulse response sample value a (t) is changed from a random variable to a fuzzy variable, and the impulse response sample value a (t) is “decreased”.
It is defined by a membership function with three states, "do it normally" and "do it big".

The control rule is IF (○○) THEN
Define in the form of (△△). For example, IF (device temperature Ta is large, error signal E is normal) THE
N (normalize step gain K, increase impulse response sample value a (t)) IF (normal device temperature Ta, large error signal E) THE
N (decrease the step gain K or increase the impulse response sample value a (t)) As described above, the states to be taken by the step gain K and the impulse response sample value a (t) that are output for all combinations of the input device temperature Ta and the error signal E are defined as control rules. This definition is made according to a rule of thumb. In addition, here (○○) below IF is the antecedent part, THE
A value equal to or less than N (ΔΔ) is called a consequent part.

By doing so, the step gain K and the impulse response sample value a (t) can be promptly determined with respect to the input device temperature Ta and the level of the error signal E. Then, the arithmetic control unit 4 calculates the filter coefficient α (t) of the virtual filter based on the determined step gain K and the impulse response sample value a (t), and based on that, the filter coefficient of the noise elimination filter 1 is calculated. Can be promptly determined and controlled.

Various modifications are possible in carrying out the present invention. For example, in information processing equipment, an intake air duct and an exhaust air duct may be provided and an active noise control system may be attached to both ducts. In this case, the systems of each system may be affected by the state of the other system. However, the adaptive filter coefficient determination method of the present invention can be applied to such a two-system system. FIG. 5 shows such a system. In each of the systems # 1 and # 2, in addition to the device temperature of the own system and the error signal, the device temperature of the partner system is also input, and the device temperature of the own system and the error signal The step gain K and impulse response sample value a of the own system for all combinations of the device temperature states of the partner system
The control rule defines the state that (t) should take.

Further, although the case where the filter coefficient determination method of the present invention is applied to the active noise control system has been described in the above embodiment, the present invention is not limited to this, and for example, various echo cancellers, noise cancellers, As an adaptive algorithm that calculates the coefficient of the applied filter in the system that uses the applied filter to generate the pseudo signal that eliminates unnecessary signals, such as ghost canceller, active vibration control system, and active suspension.
The present invention can be applied when the Filtered-X / NLMS method is used.

[0029]

As described above, according to the present invention, it is possible to quickly determine the filter coefficient of the adaptive filter by using the fuzzy theory. This enables stable adaptive control in response to load fluctuations and disturbances, and enables high-speed control for system startup and shutdown. When the present invention is applied to, for example, an active noise control system, stable noise elimination is possible even with environmental changes such as temperature fluctuations.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is a diagram illustrating a concept of an adaptive filter coefficient determination method according to an embodiment of the present invention.

FIG. 3 is a diagram showing a membership function on the input side used in the embodiment.

FIG. 4 is a diagram showing a membership function on the output side used in the embodiment.

FIG. 5 is a diagram showing another embodiment of the present invention.

FIG. 6 is a diagram showing a control system of an active noise control system.

FIG. 7 is a diagram showing an estimation system of adaptive filter coefficients by a Filtered-X / NLMS method.

FIG. 8 is a diagram showing a recursive filter expression of a Filtered-X / NLMS method.

[Explanation of symbols]

 1 Noise Elimination Filter 2 Loop Filter 3 Scattering Estimating Filter 4 Calculation Control Unit 5 Speaker 6 Noise Detection Microphone 7 Error Detection Microphone 8 Duct

Claims (2)

[Claims] 1. A method for determining a filter coefficient of an adaptive filter, which uses a Filtered-X / NLMS method in which an error scattering system is considered as an algorithm for calculating a filter coefficient, wherein a device temperature (Ta) and a coefficient calculation unit of the adaptive filter. The error signal (E) input to each is defined by the membership function, the step gain (K) and the impulse response sample value (a (t)) of the error scattering system are defined by the membership function, and the apparatus temperature And an error signal as the antecedent part, the step gain and the impulse response sample value as the consequent part are defined, and the step gain and the impulse response sample value are obtained by fuzzy inference from the device temperature and the state of the error signal, A filter coefficient determining method for an adaptive filter, which determines a filter coefficient of an adaptive filter based on the obtained result. 2. The present invention is applied to an active noise control system in which pseudo noise having the same amplitude and opposite phase as noise is generated by an adaptive filter, and the pseudo noise is used to suppress the noise. The method for determining a filter coefficient of an adaptive filter according to claim 1, which is used as an error signal.