JP3636772B2 - Fuzzy noise canceler - Google Patents

Description

【００３８】
図４を参照して，図３の実施例の動作を説明する。
サンプリングパルスジェネレータ２２のサンプリングのタイミングに従って入力信号をサンプリングし入力信号の時系列データ｛Ｘ_j ｝を生成する。グレード演算部３９は各サンプリングＸ_j のグレードを求める（図４ (b)参照）。求めたグレードは，旧グレード保持部４１に保持する。また，ファジィ推論部３４は各区間のグレードの最小値を求めファジィ推論する（図４ (c)参照）。脱ファジィ化部３５において重心演算部３５’は各推論結果を統合し，重心を求めＧ_j として出力する。
【００３９】
誤差演算部３７は誤差Ｅ_j を入力して，各区間のβ_S を求める。メンバシップ関数更新部３６は次の時刻の入力サンプリング列｛Ｘ_J+1 ｝，Ｅ_j とβ_S とにより最適メンバシップ関数を演算する。時刻ｊのとき区間Ｓのメンバシップ関数をμ_S,j としたとき_, 時刻ｊ＋１のμ_S,j+1 ＝μ_S,j ＋β_S である。但し，β_S は区間ｓのＫ（ｇ_j −Ｇ_j ）／ΣＸ_j ² である。
【００４０】
図３の構成において，白色雑音源２７はシステムを立ち上げるとき，白色雑音を未知系に加え，周波数応答特性を求めるために使用される。
【００４１】
【発明の効果】
本発明によれば，ＤＳＰ等の適応フィルタを使用することなく簡単な構成で，柔軟に能動的にノイズ消去を行うことのできるノイズ消去装置を提供することができる。
【図面の簡単な説明】
【図１】本発明の基本構成を示す図である。
【図２】本発明の基本構成の説明図である。
【図３】本発明の実施例を示す図である。
【図４】本発明の実施例の説明図である。
【図５】従来の技術を示す図である。
【符号の説明】
１：未知系
２：サンプリング部
３：ファジィ適応フィルタ
４：加算部
１１：脱ファジィ化部
１２：ファジィ推論部
１４：グレード演算部
１５：メンバシップ関数保持部
１６：メンバシップ関数更新部
１７：旧グレード保持部
１８：誤差演算部[0001]
[Industrial application fields]
The present invention relates to a fuzzy noise canceling apparatus that actively cancels noise by fuzzy inference.
[0002]
It actively eliminates noise in signal processing systems such as active noise control devices, vibration control systems, image processing systems, and voice processing systems by applying fuzzy theory.
[0003]
[Prior art]
FIG. 5 shows a system of a conventional active noise canceling apparatus. The transfer function of a system (unknown system) through which an input signal X _j propagates is estimated, and a noise canceling signal is actively generated by an adaptive filter to be unknown. This represents a system that applies noise to the output signal of the system and eliminates noise.
[0004]
In FIG.
Reference numeral 110 denotes an unknown system that transmits the input signal X _j .
An adder 111 adds noise N _j (disturbance) to the output g _j of the unknown system 110.
[0005]
Reference numeral 112 denotes an adder that adds the noise cancellation signal G _j to a signal including noise.
Reference numeral 113 denotes an adaptive filter that generates an optimum noise cancellation signal based on the input X _j and the output E _j .
[0006]
Reference numeral 114 denotes an optimum coefficient calculation unit for the adaptive filter, which calculates the optimum coefficient by the NLMS method.
The operation of the configuration of FIG. 5 will be described.
[0007]
The input signal X _j propagates through the unknown system 110 and outputs an output g _j . The output g _j of the unknown system 110 is obtained by adding the noise N _j by the adder 111 and applying the noise cancellation signal G _j by the adder 112 and taking the difference between g _j and G _j . The error E _j to which the noise cancellation signal is applied is input to the optimum coefficient calculation unit 114, and the optimum coefficient calculation unit 114 calculates the coefficient of the adaptive filter that is optimal for noise cancellation. The adaptive filter 113 generates a noise cancellation signal G _j based on the input signal X _j by the optimum coefficient calculated by the optimum coefficient calculation unit 114. However, the acoustic feedback system is omitted in FIG.
[0008]
[Problems to be solved by the invention]
In the conventional active noise canceller, the adaptive filter is constituted by an FIR filter, and the error signal (output) is controlled by a unique scalar quantity, so that adaptive control cannot be performed with flexibility. Further, since the circuit is configured using a DSP or the like, the circuit scale is increased. In particular, when the frequency response of the transmission system is increased, the number of taps of the digital FIR filter is increased and the processing speed is delayed.
[0009]
It is an object of the present invention to provide a fuzzy noise canceling apparatus capable of actively performing noise canceling with a simple configuration without using an adaptive filter such as a DSP.
[0010]
[Means for Solving the Problems]
The fuzzy noise canceling apparatus of the present invention is a noise canceling apparatus that cancels noise by applying a noise canceling signal to the output of an unknown system whose transfer function is unknown. A means for sampling by dividing into intervals, and a membership function corresponding to the sampling data for each interval are determined, and updating of the membership function is repeated for each sampling. A means for actively updating the membership function so that the error in each sampling interval is minimized by the normalized least square method based on the error of the previous sampling period in the sampling period, and the input signal And the membership function to estimate the transfer function of the unknown system and actively generate the optimal noise cancellation signal. And to have a fuzzy adaptive filter formed.
[0011]
FIG. 1 is a diagram showing a basic configuration of the present invention.
In FIG.
1 is an unknown system.
[0012]
Reference numeral 2 denotes a sampling unit.
Reference numeral 3 denotes a fuzzy adaptive filter.
Reference numeral 4 denotes an adder that adds the noise cancellation signal G _j output from the fuzzy adaptive filter 3 to the unknown signal output and takes the difference from the unknown output g _j .
[0013]
In the fuzzy adaptive filter 3,
Reference numeral 11 denotes a defuzzification unit.
Reference numeral 12 denotes a fuzzy reasoning unit.
[0014]
Reference numeral 14 denotes a grade calculation unit for obtaining a grade by a membership function with the input sampling value X _j as a support.
A membership function holding unit 15 holds a membership function.
[0015]
Reference numeral 16 denotes a membership function update unit for updating the membership function. The membership function is updated based on, for example, the method of least squares and follows μ _{S, j + 1} = μ _{S, j} + KX _j (g _j −G _j ) / ΣX _j ² . Here, s represents a section obtained by equally dividing time-series n sampling data of the input signal into N equal parts. Σ is the sum for each j in the interval s.
[0016]
Reference numeral 17 denotes an old grade holding unit which holds the grade calculated by the grade calculating unit 14 as an old grade (μ _j ).
Reference numeral 18 denotes an error calculation unit. An error is input, and an optimum value necessary for updating the membership function by, for example, the NLMS method is obtained, and KX _j (g _j −G _j ) / ΣX _j ² is calculated for each section. Hereinafter, KX _j (g _j −G _j ) / ΣX _j ² = β, and β in the section _S is represented by β _S.
[0017]
[Action]
The operation of the basic configuration of the present invention shown in FIG. 1 will be described with reference to FIG.
FIG. 2A shows time series data {X _j } of the sampling X _j of the input signal, and S ₁ , S ₂ ,..., S _N are sections obtained by dividing the input sampling sequence X _j at a predetermined period. Yes, membership functions are assigned to each section. For example, S ₁ is a membership function M ₁ , S ₂ is a membership function M ₂ , and S _N is a membership function M _N.
[0018]
FIG. 2B shows membership functions (M ₁ , M ₂ ,..., M _N ), which are assigned to the input sampling intervals. For example, when fuzzy inference is performed by the MIN-MAX method, a grade is obtained using the sampling value of each section as a support, and the minimum value is adopted for inference. In FIG. 2 (b), μ ₁ , μ ₂ ,..., Μ _N are the minimum values of the grades thus determined. That is, μ ₁ is the minimum grade value when the membership function M ₁ is applied to the sampling data in the section S ₁ . In the case of FIG. 2 (b), the minimum value of the grade in the section of S ₁ is the grade of support V ₂ (corresponding to the sampling value of j = 2). μ ₂ is the minimum value of the grade when the membership function M ₂ is applied to the sampling data in the section S ₂ . μ _N is the minimum value of the grade when the membership function M _N is applied to the sampling data in the section S _N.
[0019]
Fig. 2 (c) shows the inference results obtained in this way.
Fig. 2 (d) is an integration of the inference results of Fig. 2 (c). G _j is the center of gravity of the integration result.
[0020]
Assume that the input sequence {X _j } of the sampling signal in FIG. 2A is at time t _j . As shown in FIG. 2 (b), the grade calculating unit 14 determines each X _j (for each section based on the membership function (M ₁ , M ₂ ,..., M _N ) for each section and the input signal X _j. The grade of j = 0-4) is obtained. The obtained grade is held in the old grade holding unit 17.
[0021]
The fuzzy reasoning unit 12 performs fuzzy reasoning for each section based on, for example, the minimum value based on the grade for each section (see FIG. 2C). The defuzzification unit 11 integrates the inference results of each section, and obtains and outputs G _j from the center of gravity, for example.
[0022]
On the other hand, the error calculation unit 18 obtains β (= KX _j (g _j −G _j ) / ΣX _j ² ) based on the error E _j . Σ is the sum for j in each interval. That is, for j = 0 to 4 in the section S _1, obtains a beta, similarly section S _2, · · ·, for S _N beta _2, · · ·, seek beta _N. Then, the membership function updating unit 16 obtains the membership function by μ _{S, j + 1} = μ _{S, j} + β _{S according} to β _S and old grade for each section obtained by the error calculation unit 18. It is held by the holding unit 15.
[0023]
Based on the input sampling sequence at the next time _{t j + 1 {X j +} 1} performs the operation described above, obtains and outputs the noise cancellation signal G _{j + 1} at time t _{j + 1.}
According to the present invention, it is possible to provide a noise canceling apparatus that can perform noise canceling flexibly and actively with a simple configuration without using an adaptive filter such as a DSP.
[0024]
【Example】
FIG. 3 shows an embodiment of the present invention.
In FIG.
21 is an unknown system that transmits an input signal.
[0025]
A sampling pulse generator 22 samples the input signal.
A multiplication unit 23 multiplies the pulse of the sampling pulse generator 22 and the input signal to generate a sampling input signal _Xj .
[0026]
An adder 24 adds noise _Nj to the output of the unknown system.
An adder 25 adds a noise cancellation signal G _j to an unknown output including noise.
[0027]
Reference numeral 26 denotes a fuzzy adaptive filter.
Reference numeral 27 denotes a white noise source (M-seq).
In the fuzzy adaptive filter 26,
Reference numeral 34 denotes a fuzzy reasoning unit.
[0028]
Reference numeral 34 'denotes a minimum value judging unit for obtaining the minimum value of the grade in each section.
Reference numeral 35 denotes a defuzzification unit.
[0029]
Reference numeral 35 'denotes a center of gravity calculation unit for obtaining the center of gravity by integrating the results of fuzzy inference of each section.
Reference numeral 36 denotes a membership function update unit which updates the membership function according to μ _{S, j + 1} = μ _{S, j} + KX _j (g _j −G _j ) / ΣX _j ² . s represents a section, and Σ represents the sum of j in each section. For example, an average of each discrete section of j = J + 1 to (n + 1) J is taken.
[0030]
An error calculation unit 37 calculates β (= KX _j (g _j −G _j ) / ΣX _j ² ) necessary for obtaining an optimum filter coefficient by NLMS.
Reference numeral 38 denotes a K holding unit which holds a step gain K.
[0031]
Reference numeral 39 denotes a grade calculation unit.
Reference numeral 40 denotes a membership function holding unit 40.
Reference numeral 41 denotes an old grade holding unit.
[0032]
FIG. 4 is a diagram for explaining the operation of the embodiment of the present invention.
FIG. 4A shows a time-series signal of sampling data of the input signal, the sampling time interval is 250 μs, and the time is divided every 125 μs (every 5 samplings). _{t 1, t 2, ··· t} 8 is separated in time, the separator of each time _{S 1, S 2, S 3} , ···, and S _8. FIG. 4 (a) represents 100 Hz signal sampling.
[0033]
Figure 4 (b) shows the membership function at each time interval. M ₁ is a membership function in S ₁ , M ₂ is a membership function in S ₂ , M ₃ is a membership function in S ₃ , and M ₈ is a membership function in S ₈ . The horizontal axis (support) of each membership function is a sampling value in each section. In M ₁ , a ₀ , a ₁ , a ₂ , a ₃ , and a ₄ are sampling values at j = ₀ , ₁ , ₂ , ₃ , and ₄ , respectively, and the grades are μ _1,0 , μ _{1, 1} , μ _1,2 , μ _1,3 , μ _1,4 . Similarly, in M ₂ , b ₀ , b ₁ , b ₂ , b ₃ , and b ₄ are sampling values at j = 5, 6, 7, 8, and 9, respectively, and each grade is μ _2,0 , μ _2,1 , μ _2,2 , μ _2,3 , μ _2,4 . In M ₃ , d ₀ , d ₁ , d ₂ , d ₃ , and d ₄ are sampling values at j = 10, 11, 12, 13, and 14, respectively, and their grades are μ _3,0 , μ _{3. , 1} , μ _3,2 , μ _3,3 , μ _3,4 . In M ₈ , h ₀ , h ₁ , h ₂ , h ₃ , and h ₄ are sampling values at j = 35, 36, 37, 38, and ₃₉ , and the grades are μ _8,0 , μ _8, respectively _{. 1,} μ _8,2, μ _8,3, it is μ _8,4.
[0034]
In the fuzzy inference unit 34, the minimum value determination unit 34 ′ obtains the minimum value of the grade in each section. Section _{S 1, S 2, S 3} , S 4, ···, the minimum grade in _{S 8 μ 1, μ 2,} μ 3, ···, and mu _8. FIG. 4 (c) shows the result of fuzzy inference by taking MAX of each membership function and support (average value in sampling data interval) in FIG. 4 (b).
[0035]
In the above, the membership function update is
_{μs, (j + 1)} = _{μs, j} + β _S (β _S is KX _j (g _j −G _j ) / ΣX _j ² in section S).
[0036]
The center of gravity G _j is calculated by the following formula.
[0037]
[Expression 1]

[0038]
The operation of the embodiment of FIG. 3 will be described with reference to FIG.
The input signal is sampled according to the sampling timing of the sampling pulse generator 22 to generate time series data {X _j } of the input signal. The grade calculating unit 39 obtains the grade of each sampling X _j (see FIG. 4B). The obtained grade is held in the old grade holding unit 41. Further, the fuzzy reasoning unit 34 obtains the minimum grade value of each section and performs fuzzy reasoning (see FIG. 4C). In the defuzzification unit 35, the center-of-gravity calculation unit 35 ′ integrates the respective inference results, obtains the center of gravity, and outputs it as G _j .
[0039]
The error calculator 37 receives the error E _j and obtains β _S for each section. The membership function updating unit 36 calculates an optimal membership function from the input sampling sequence {X _{J + 1} }, E _j and β _{S at the} next time. When the membership functions of the section S at time j was mu _{S, j,} a time j + 1 of _{μ S, j + 1 = μ} S, j + β S. However, β _S is K (g _j −G _j ) / ΣX _j ² in the section s.
[0040]
In the configuration of FIG. 3, the white noise source 27 is used to add the white noise to the unknown system and obtain the frequency response characteristics when starting up the system.
[0041]
【The invention's effect】
According to the present invention, it is possible to provide a noise canceling apparatus that can perform noise canceling flexibly and actively with a simple configuration without using an adaptive filter such as a DSP.
[Brief description of the drawings]
FIG. 1 is a diagram showing a basic configuration of the present invention.
FIG. 2 is an explanatory diagram of a basic configuration of the present invention.
FIG. 3 is a diagram showing an embodiment of the present invention.
FIG. 4 is an explanatory diagram of an embodiment of the present invention.
FIG. 5 is a diagram showing a conventional technique.
[Explanation of symbols]
1: unknown system 2: sampling unit 3: fuzzy adaptive filter 4: addition unit 11: defuzzification unit 12: fuzzy inference unit 14: grade calculation unit 15: membership function holding unit 16: membership function update unit 17: old Grade holder 18: Error calculator

Claims (5)

Ri by that transfer function applies the noise cancellation signal to the output of the unknown unknown system, a noise erasing apparatus for erasing noise,
Means for sampling an unknown input signal by dividing it into equal intervals in time series ;
A membership function is determined corresponding to the sampling data for each section, and the membership function is updated every sampling. The error between the output of the unknown system and the noise cancellation signal is the previous sampling in the sampling period. Means for actively updating the membership function so that the error in each sampling interval is minimized by the normalized least square method based on the error of the period ;
A fuzzy noise canceling device comprising a fuzzy adaptive filter that performs fuzzy inference based on an input signal and a membership function, estimates a transfer function of an unknown system, and actively generates an optimal noise canceling signal .   2. The fuzzy noise canceling apparatus according to claim 1, further comprising: a white noise source, wherein white noise is input to an unknown system, and the frequency response characteristic of a system including the unknown system is obtained by fuzzy inference.   3. The fuzzy inference is performed by sampling data in the section and a membership function of the section, fuzzy inference results in each section are integrated, and a noise cancellation signal is output based on the integration result. Fuzzy noise elimination device. When the grade based on X _j of the sampling signal at time t _j in the interval s and μ _{S, J,} of the next cycle interval s time t _{j + 1} of X _{j + 1} grade based on mu _{S, j + 1} = μ _{S, j} + KX _j E _j / ΣX _j ² (where K is a step gain, E _j is an error, and Σ is a sum of j in each interval). The fuzzy noise erasing apparatus according to 1, 2, or 3. Σ of μ _{S, j + 1} = μ _{S, j} + KX _j E _j / ΣX _j ^{2 in} the normalized least square algorithm 5. The fuzzy noise erasing apparatus according to claim 4, wherein the fuzzy noise erasing apparatus is performed by averaging in a section.

Images