JPH0990962A - Active muffling device - Google Patents

Abstract

(57) [Abstract] [Problem] As a result of analyzing noise that was considered to be a random sound, a point that is chaotic is found, and the noise ahead is predicted in time from the noise signal to achieve compactness. A predictor (50) that predicts a noise (11) that precedes a noise signal in time and outputs a predicted noise signal is provided. The FIR filter (41) outputs an inverted sound signal having the same phase and an opposite phase with respect to the predicted noise signal, and the speaker (31) emits the inverted sound. The monitor microphone (32) outputs the error between the inverted sound and the noise (11), and updates the filter coefficient of the FIR filter (41) so that the error becomes small. The predictor (50) includes a data processing circuit (51) for outputting a predicted noise signal one sampling ahead based on a neural network in which the number of input layer units is determined based on the interphase dimension of chaos determination, and the predicted noise signal. And a coefficient calculation circuit (52) for calculating a weighting coefficient of the neural network based on the noise signal.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an active muffling device, and more particularly to measures for downsizing.

[0002]

2. Description of the Related Art Conventionally, as disclosed in Japanese Unexamined Patent Publication No. 5-323974, an active silencer of this type is provided with a detection microphone, a speaker and a monitor microphone installed in a duct from the upstream side of noise. The FIR filter generates an inverted sound signal based on the noise detected by the microphone, and the speaker emits the inverted sound to the duct based on the inverted sound signal to cancel the noise.

On the other hand, the monitor microphone detects an error between the noise and the inverted sound, and the adaptive algorithm execution circuit updates the filter coefficient of the FIR filter so that the error becomes small based on the error signal of the monitor microphone. I have to.

[0004]

In the above-mentioned conventional active silencer, it is desired to make the entire device compact in order to increase the degree of freedom of installation.

On the other hand, in the above-described active silencer, there is a delay in the acoustic equipment such as the detection microphone and the speaker and the controller itself which processes the analog signal and the digital signal. It requires a certain distance from the speaker.

In order to achieve the above-mentioned compactness, it is necessary to shorten the distance between the detection microphone and the speaker, but in order to shorten this distance, it is necessary to reduce the system delay as much as possible.

In that case, the delay of the digital part of the controller is improved by increasing the sampling frequency due to the progress of the control chip. However, since the delay due to the audio equipment is a physical problem, there is a problem that it cannot be expected to eliminate the delay.

Therefore, the conventional active silencer described above has a limit in its compactness.

The present invention has been made in view of the above point, and the inventors of the present invention have found out that it is chaotic as a result of analyzing noise that was considered to be a random sound, and this new fact The purpose of the present invention is to predict the noise ahead in time from the noise signal based on the above, and to emit a reverse sound to achieve compactness.

[0010]

In order to achieve the above object, as shown in FIG. 1, the means taken by the invention according to claim 1 is such that noise (11) propagating in a propagation space (21) ) Is detected and a noise signal (30) for outputting a noise signal is provided. Then, the noise signal output from the noise detection means (30) is received and the noise (1
Prediction means (50) that predicts 1) and outputs a predicted noise signal
Is provided. Further, a digital filter (41) for generating and outputting an inverted sound signal having the same amplitude and an opposite phase with respect to the predicted noise signal output by the prediction means (50), and an inverted signal of the digital filter (41). Reverse sound emitting means (31) for receiving the reverse sound and radiating the reverse sound to the propagation space (21) is provided. In addition, an error detection means (32) for detecting an error between the inversion sound emitted by the inversion sound emission means (31) and the noise (11) in the propagation space (21) and outputting an error signal, and the error detection means. An adaptive algorithm execution circuit (44) for updating the filter coefficient of the digital filter (41) so as to reduce the error by receiving the error signal output by the means (32) and the predicted noise signal output by the prediction means (50). And are provided.

The means taken by the invention according to claim 2 is that in the invention according to claim 1, the predicting means (50) is
The noise (11) at one sampling destination is predicted and the predicted noise signal is output.

Further, the means taken by the invention according to claim 3 is that in the invention according to claim 2, the predicting means (50) is
A data processing circuit (51) which inputs a noise signal and outputs a predicted noise signal based on a neural network, and a weight of each unit in the neural network based on a difference between the predicted noise signal and an actual noise signal corresponding to the predicted noise signal A coefficient calculation circuit (52) for calculating a coefficient and outputting a coefficient signal to the data processing circuit (51) is provided.

Further, the means taken by the invention according to claim 4 is the data processing circuit (5
In 1), the number of units in the input layer is determined based on the interphase dimension that determines the chaoticity of noise signals.

-Operation- Due to the above-mentioned matters specifying the invention, in the invention according to claim 1,
First, the noise (11) propagating in the propagation space (21) is detected by the noise detecting means (30) and input to the predictor (50).

The predictor (50) predicts the noise (11) preceding the noise signal in time and outputs the predicted noise signal. Specifically, in the inventions according to claims 2 and 3, the data processing circuit (51) acquires the noise signal from the noise detecting means (30) and outputs the predicted noise signal. On the other hand, an erroneous calculation signal obtained by comparing the predicted noise signal delayed by one sampling time with the noise signal is input to the coefficient calculation circuit (52).

Then, the coefficient calculation circuit (52) calculates a weighting coefficient based on the erroneous calculation signal and outputs the coefficient signal to the data processing circuit (51). Then, the data processing circuit (51) is
The prediction noise signal at the sampling destination is output, and the coefficient calculation circuit (52) updates the weighting coefficient of the neural network. In that case, in the invention according to claim 4, the number of units in the input layer of the neural network is determined based on the interphase dimension of chaos.

On the other hand, the predicted noise signal is input to the digital filter (41) and the adaptive algorithm execution circuit (44). Then, the digital filter (41) generates an inverted sound signal having a phase opposite to that of the predicted noise signal and having the same amplitude. This inverted sound signal is generated by the inverted sound emitting means (3
The reverse sound is output to 1) and is emitted to the propagation space (21). Due to this radiation, the noise (11) in the propagation space (21)
Good reduction due to inversion sound.

Further, the noise level attenuated by the inversion sound, that is, the error is detected by the error detecting means (32), and the error signal is input to the adaptive algorithm execution circuit (44). The filter coefficient of the digital filter (41) is sequentially updated by the adaptive algorithm execution circuit (44) based on the error signal and the predicted noise signal. By updating this filter coefficient, the inverted sound signal is generated by the digital filter (41) with respect to the noise (11) with high accuracy over time,
Noise (11) is best reduced.

[0019]

Therefore, according to the first aspect of the present invention,
Since the noise signal is predicted, the distance between the noise detecting means (30) and the reversal sound emitting means (31) can be shortened by the amount corresponding to the prediction time. Can be planned. As a result, the degree of freedom of installation can be improved and the range of application can be expanded.

Further, according to the inventions according to claims 2 and 3, the predicting means (50) has a noise (1
Since the above 1) is predicted, the data processing circuit (51) for updating the weighting factor can be omitted, so that the configuration of the predictor (50) can be simplified.

According to the invention of claim 4, since the number of units of the input layer in the neural network is set based on the interphase dimension for determining the chaos of the time series data, the deterministic relationship is determined by the neural network. It becomes easy to express in the network, and the accuracy of the predicted noise signal can be improved. This results in noise (11)
The attenuation level of can be improved.

[0022]

Embodiment 1 of the present invention will be described below in detail with reference to the drawings.

As shown in FIG. 2, the active silencer (10) silences the noise (11) in the air conditioning duct (20), and the inside of the duct (20) of FIG. A propagation space (21) in which noise (11) propagates from a noise source (not shown) located at the left end is formed.

In the duct (20), the detection microphone (30) and the speaker (31) are located from the upstream side to the downstream side in the propagation direction of the noise (11) in the propagation space (21) shown by the arrow in FIG. And a monitor microphone (32) are arranged in that order. The detection microphone (30) constitutes noise detection means for detecting the noise (11) in the propagation space (21), and a noise signal is controlled by a controller (40) through a low-pass filter (not shown) for preventing aliasing. Output to.

The speaker (31) is connected to the controller (4
0) for receiving a reversal sound signal from a low-pass filter (not shown) for preventing aliasing, and for radiating a reversal sound having a phase opposite to that of the noise (11) and the same amplitude to the propagation space (21). It is an additional sound source and constitutes an inverted sound emitting means.

The monitor microphone (32) is arranged at a predetermined observation point and detects at the observation point the reduced sound level of the noise (11) reduced by the action of the reversal sound emitted from the speaker (31). In addition, an error detection unit that outputs an error signal via a low-pass filter (not shown) for preventing aliasing is configured. That is, the monitor microphone (32) detects an error between the inverted sound from the speaker (31) and the noise (11) and outputs an error signal.

The controller (40) is for generating an inverted sound signal having a phase opposite to that of the noise (11) and having the same amplitude. The inverted sound signal to the speaker (31) is a D / A converter. (Not shown) and the like, and at the same time, the noise signal of the detection microphone (30) and the error signal of the monitor microphone (32) are input via the A / D converter (not shown) and the like. .

The controller (40) includes an adaptive FIR filter (41), a first correction filter (42) and a second correction filter (43), and an adaptive algorithm execution circuit (44). The noise signal output from the detection microphone (30) is input to the predictor (50) via the adder circuit (45).

The adaptive FIR filter (41) receives a predictive noise signal output from a predictor (50) described later, and basically generates a reversal sound signal having the same amplitude and a reverse phase to the predictive noise signal. It constitutes a digital filter.

Further, in the first correction filter (42), there is an error path (3c) in which the inverted sound radiated from the speaker (31) is input to the monitor microphone (32).
The predictive noise signal of the predictor (50) described later is received, and the predictive noise signal is corrected with respect to the transfer function from the output of the inverted sound to the error path (3c). Specifically, the first correction filter (42) delays the predicted noise signal in advance for a predetermined time by the propagation time required for the inverted sound signal to be input from the speaker (31) to the monitor microphone (32), and also attenuates the waveform. The filter is a FIR filter having a transfer function between the speaker (31) and the monitor microphone (32).

The second correction filter (43) detects the inverted sound radiated from the speaker (31) and detects the inverted sound (3).
Since there is a feedback path (3b) propagating to (0), it is a filter for preventing howling from occurring due to this feedback path (3b). Specifically,
The second correction filter (43) is an adaptive FIR filter (4
The inverted sound signal of 1) is filtered and fed back to the addition circuit (45), and is composed of an FIR filter having a transfer function from the output of the inverted sound to the feedback path (3b).

The adaptive algorithm execution circuit (44) is
It is configured to perform adaptive control by a least mean square method (LMS; Least Mean Square) algorithm. The adaptive algorithm execution circuit (44) is a signal obtained by delaying the predicted noise signal received through the first correction filter (42),
Based on the error signal output from the monitor microphone (32), the filter coefficient of the adaptive FIR filter (41) is updated by LMS so that the error between the noise signal and the inverted sound is reduced, and the inverted sound signal is corrected. Adaptive control.

On the other hand, the predictor (50), which is a feature of the present invention, constitutes a predicting means for predicting and outputting a noise signal at one sampling destination, and as shown in FIG. A circuit (51) and a coefficient calculation circuit (52) are provided.

The data processing circuit (51) receives the noise signal from the detection microphone (30) via the adding circuit (45) and processes the time-series data which is the noise signal based on a neural network. The noise signal at one sampling destination is temporally predicted with respect to the currently sampled noise (11), and the predicted noise signal that is time-series prediction data is output.

The coefficient calculation circuit (52) is a circuit for calculating the weighting coefficient of the neural network in the data processing circuit (51), and the error signal between the predicted noise signal and the noise signal is input from the comparison circuit (53). Has been done. That is,
The predicted noise signal output from the data processing circuit (51) corresponds to the noise signal at one sampling destination. Therefore, the predicted noise signal is delayed by one sampling by the delay circuit (54), and then the detection microphone (30 The noise signal from () is compared by the comparison circuit (53) and an error signal is output. The coefficient calculation circuit (52) receives the error signal, calculates the weighting coefficient, and outputs the coefficient signal to the data processing circuit (51). Then, the data processing circuit (51) adjusts the weighting coefficient and outputs the predicted noise signal.

-Prediction Principle- Therefore, the basic principle for predicting the noise (11) will be described.

First, in predicting the noise (11), the frequency characteristic of the noise (11) propagating through the duct (20), that is, the fan noise (11) is analyzed, and the result is shown in FIG. That is, as is clear from FIG. 7, it can be seen that the fan noise (11) contains a lot of low-frequency noise (11) of 500 Hz or less.

Since there are many low frequency components, it can be said that the time variation with respect to the phase of the noise (11) is gentle. Therefore, accurate prediction is possible in a short time.

On the other hand, the above-mentioned neural network is effective for pattern recognition, as represented by the use in image recognition, and time-series data containing many low-frequency components causes pattern fluctuations in a short time. Since it is small, it can be expected that the neural network is used to predict the time series data with relatively high accuracy based on the pattern recognition characteristics. Therefore, it can be said that the fan noise, which is the noise (11), can be predicted by the neural network.

However, the time-series data of the fan noise (11) exhibits the phenomenon shown in FIG. 8, and the fan noise (11) is
Since no property can be found from FIG. 8, it was conventionally considered as a random sound.

As a result of earnest research on the noise (11), the inventors of the present invention newly found that the low-frequency fan noise (11) was chaotic, and completed the present invention. Therefore, the consideration regarding this chaos will be explained.

First, the reconstructed trajectory will be described.
To investigate the chaotic behavior of time series data, consider a two-dimensional plane embedding space. That is, regarding the time-series data of the noise (11), the data X (t) at a certain time and the data X (t + T) temporally advanced from the data X (t) by a predetermined time T (delay time). And take this X (t) and X (t + T)
Considering a two-dimensional plane (embedded space) having two and as two axes, the trajectories of points | X (t) and X (t + T) | with the change of time t are as shown in FIG. If the noise (11) in FIG. 8 is completely random, the points | X (t) and X (t + T) | will be dispersed, but as is clear from FIG.
The orbits of (t) and X (t + T) | are not dispersed and are limited to a predetermined range, and are considered chaotic.

Second, the Lyapunov exponent will be described.
The Lyapunov exponent is a quantity that represents the degree to which orbits close to the reference orbit become separated with time in the phase space, and chaoticity can be determined by the positive or negative of the Lyapunov exponent. As shown in FIG. 10, the Lyapunov exponent λ uses the Wolf algorithm and the reference point Xi (t) at the time t in the phase space and the point Yi (t) closest to the reference point Xi (t). Then, by calculating the ratio of the distances of τ intervals while following the time change of these two points and averaging over many pairs | Xi (t), Yi (t) |
It is expressed by the following equation (1). It should be noted that ‖… ‖ in the following equation (1)
Represents the length of the vector.

[0044]

(Equation 1)

When this Lyapunov exponent λ is smaller than zero (λ ≦ 0), the amount of separation is zero or small, which indicates that the motion is a point attractor or periodic attractor, and the Lyapunov exponent λ is smaller than zero. When larger (λ>
0), it has chaos. If the Lyapunov exponent λ is considerably larger than zero, the amount of separation is too large, indicating a random phenomenon.

The evaluation of the Lyapunov exponent λ for the time series data of the noise (11) shown in FIG. 8 is as shown in FIG. As is clear from FIG. 11, the Lyapunov exponent λ for the time series data of the noise (11) is about 0.003.
It can be said that the above noise (11) is chaotic.

Third, the interphase dimension will be described. A fractal dimension is a quantity that characterizes a strange attractor in a phase space. This fractal dimension is a dimension that is extended to a non-integer value (decimal value), unlike a topological dimension that handles only integer values, which is called a normal dimension. One method for obtaining this fractal dimension is to use Grossberger (P. Grassber
ger) and Procaccia (I. Procaccia), and the correlation dimension is obtained by the following equation (2) using the correlation integral C (r). In the following equation (2), ‖ ... ‖ represents the length of the vector.

[0048]

(Equation 2)

H (α) in the above equation (2) is a Heaviside function, and is represented by the following equation (3).

[0050]

(Equation 3)

It is considered that the correlation integral C (r) calculated by the above equation (2) can obtain the relationship shown in the following equation (4) within an appropriate range of the radius r.

[0052]

(Equation 4)

Here, ν (D) is an interphase index. This phase index ν (D) is the radius logr on the horizontal axis and the correlation integral l on the vertical axis.
In the graph plotted by taking ogC (r), it is determined by the slope of the straight line portion within the range of an appropriate radius r.

Then, the correlation integral C (r) is calculated while successively increasing the embedding dimension D, and the interphase exponent ν (D) is obtained.
The interphase exponent ν (D) increases until it exceeds the dimension D required for embedding the actual attractor, but eventually saturates. That is, the asymptotic value Dcorr becomes the interphase dimension, and if the value of the interphase dimension Dcorr is a non-integer value including a small number, it can be said to have chaotic properties.

Therefore, with respect to the time series data of the fan noise (11) shown in FIG. 8, the correlation integral log with respect to the radius logr
When C (r) is obtained, it becomes as shown in FIG. 12, and when the interphase index ν (D) for the embedding dimension D is obtained, it becomes as shown in FIG. From this FIG. 13, the interphase dimension Dcorr of the fan noise (11) is a non-integer value of 3 and 4.

From the above results, it can be determined that the fan noise (11) is deterministic chaos. If the fan noise (11) has chaoticity, it is considered that there is some deterministic relationship. Therefore, if this deterministic relationship can be expressed by a neural network, noise (1
It is possible to predict the time series data of 1).

In order to reproduce this deterministic relationship, FIG.
Therefore, it is necessary to consider the dimension which is about twice as large as the interphase dimension Dcorr. That is, since the embedding dimension D of the interphase dimension Dcorr in FIG. 13 converges to a value of 6 to 7, it is necessary to consider a space of 6 to 7 dimensions.

For example, when the attractor is a straight line and is one-dimensional, if the embedding dimension D, which is the dimension of the phase space, is successively increased from 1, the interphase exponent ν (D) saturates in one dimension.
Even if the embedding dimension D is increased to 1 or more, the interphase index ν (D) is one-dimensional and the interphase dimension Dcorr is one-dimensional. If the attractor is a torus with three dimensions and the embedding dimension D, which is the dimension of the phase space, is successively increased from 1, the interphase exponent ν (D) becomes saturated in three dimensions, and the embedding dimension D becomes 3 or more. Even if it is increased, the correlation index ν (D) is three-dimensional and the correlation dimension Dcorr
Is three-dimensional.

Therefore, in the case of the fan noise (11), it is considered that the deterministic relationship can be reproduced in the 6 to 7-dimensional phase space, so the number of units in the input layer of the neural network is set to 6 or more. It is considered desirable to do this.

-Principle of Application of Neural Network- Next, the reason why the neural network is applied to the predictor (50) will be described with reference to FIGS. The same parts as those in the present embodiment are indicated by the same reference numerals.

First, FIG. 20 shows a conventional adaptive filter (4
The system configuration of the active muffler (10) using A) is shown. The adaptive filter (4A) receives the noise signal of the detection microphone (30) and generates an inverted sound signal, and the speaker (31) inverts it. While radiating sound, the monitor micro fan outputs noise (11), inverted sound and error signal to the adaptive filter (4A). Since there is a feedback path (3b) that propagates to the detection microphone (30) in the reversal sound radiated from the speaker (31), correction of the transfer function from the reversal sound output to the feedback path (3b) is performed. In order to perform the above, the second correction filter (43) adds the correction signal based on the inverted sound signal to the noise signal.

As shown in FIG. 21, the adaptive filter (4A) is an adaptive FI that receives an input signal which is a noise signal.
An R filter (41) and an adaptive algorithm execution circuit (44) are provided, and an error signal between the inverted sound signal and a desired noise signal is input to the adaptive algorithm execution circuit (44). Based on this error signal, the adaptive FIR filter (4
Adaptive filter algorithm execution circuit (44)
Has been updated.

Since there is an error path (3c) detected by the monitor microphone in the above-mentioned reversed sound, in order to operate the adaptive filter (4A) accurately, as shown in FIG. It is necessary to provide a filter (F1) having the characteristic of the inverse function C ⁻¹ of the transfer function in order to cancel the existence of the transfer function C from to the error path (3c).

However, since it is difficult to obtain the inverse function C ^-1 because it is necessary to make a prediction in terms of time, a filter (F2) corresponding to the transfer function C is used as shown in FIG. Since it is installed on the input side of the adaptive algorithm execution circuit (44), as shown in FIG. 24, the conventional active silencer (10) includes the first correction filter (42).
Is provided on the input side of the adaptive algorithm execution circuit (44).

Therefore, as described above, it is possible to apply a neural network to the adaptive FIR filter (41) when predicting the noise (11). However, the data processing circuit (51) to which the neural network is applied cannot be directly applied in place of the adaptive FIR filter (41).

That is, the backpropagation algorithm is used to update the weight coefficient of the neural network, but the Filtered-X algorithm cannot be applied to the update of the weight coefficient by the backpropagation method.

The reason is that the Filtered-X algorithm is
As described above, it is for correcting the existence of the transfer function C from the inverted sound output of the speaker (31) to the error path (3c), and the filter coefficient is calculated by the least mean square method (LMS). By providing the first correction filter (42), which is the filter (F2) related to the transfer function C, on the input side, the influence of the transfer characteristic can be eliminated.

However, the backpropagation algorithm used for the neural network has a large number of layers, and it is difficult to apply it because it is necessary to correct the transfer function C for each layer.

From the above, according to the present invention, the predictor (50) is provided on the input side of the adaptive FIR filter (41) and the adaptive algorithm execution circuit (44) as described above.

Here, a learning method by the back propagation algorithm in the neural network will be described.

As shown in FIG. 14, the total sum of inputs to the i-th unit in the m-th layer is expressed by the following equation (5).

[0072]

[Formula 5]

The output of the i-th unit in the m-th layer is represented by the following equation (6) by the non-linear transformation F.

[0074]

[Formula 6]

Assuming that the expected output is di (i = 1, ..., NM), the total sum of squared errors E of the net output is defined by the following equation (7).

[0076]

[Formula 7]

Considering coefficient updating by the steepest descent method with respect to the squared error E, the coefficient updating equation is as shown in the following expression (8).

[0078]

(Equation 8)

Here, η is a parameter that determines the amount of coefficient update once. In order to obtain the partial differential in the above equation (8), first, the partial differential with respect to y in the above equation (7) is performed.
The following equation (9) is obtained.

[0080]

[Formula 9]

The partial differential of the squared error E with respect to x is given by the following equation (10).

[0082]

(Equation 10)

Since the partial differential of y with respect to x in the above equation (10) is the derivative of the non-linear function, if the sigmoid function of the following equation (11) is defined as the non-linear function, the equation (1
The derivative of 2) is obtained.

[0084]

[Formula 11]

From the above equations (9), (10) and (12), the following equation (13) is obtained.

[0086]

(Equation 12)

Here, the following equation (14) is obtained from the above equation (1).

[0088]

(Equation 13)

Therefore, the target partial differentiation of the squared error E with respect to W is obtained from the above equations (13) and (14) as the following equation (15).

[0090]

[Formula 14]

This equation (15) can be calculated using the unit output calculated for a given input and the expected output given in advance, so that a heavy update to the M-th layer can be realized. It will be possible.

On the other hand, the weight update of the intermediate layer will be examined. First, consideration will be given to the coefficient update regarding the weight from the M-2 layer to the M-1 layer. In this case, since the expected output is unknown, the equation corresponding to the above equation (9) cannot be obtained. Therefore, the partial differential of the squared error E with respect to W is expressed by the following equation (16).
It becomes a relationship. However, in order to avoid ambiguity, subscripts for distinguishing the units of the Mth layer, the M-1st layer, and the M-2nd layer are shown by i, j, and k, respectively.

[0093]

(Equation 15)

The above equation (16) can be easily calculated by using the result of the equation (13).
It is possible to update the coefficient related to the weight from the layer to the (M-1) th layer. After that, the partial differential of the weights can be calculated sequentially in the direction of the input layer by the same procedure, and all the coefficients can be updated.

Data processing circuit (51) by the method described above.
However, in the present embodiment, the number of units in the input layer is set to 6 or more, while the number of units in the output layer only outputs the predicted noise signal, as described above. Therefore, it becomes one.

-Silencer Operation- Next, the silencing operation of the active silencer (10) will be described.

First, the noise (11) propagating in the propagation space (21) in the duct (20) is detected by the detection microphone (30), and the noise signal detected by this microphone (30) is added by the addition circuit ( It is input to the predictor (50) via 45). The predictor (50) operates as described below to output a predicted noise signal, and the predicted noise signal is input to the adaptive FIR filter (41).

The predicted noise signal is delayed by the first correction filter (42) for a predetermined time or the like, and then input to the adaptive algorithm execution circuit (44). on the other hand,
In the adaptive FIR filter (41), an inverted sound signal having a phase opposite to that of the predicted noise signal and the same amplitude is generated. This inverted sound signal is output to the speaker (31), and the inverted sound is radiated from the speaker (31) to the propagation space (21). Due to this radiation, the noise (11) in the duct (20) is satisfactorily reduced by the reversal sound radiated from the speaker (31).

Further, the noise level attenuated by the inversion sound, that is, the error is detected by the monitor microphone (32), and the error signal is input to the adaptive algorithm execution circuit (44). Based on this error signal and the predicted noise signal from the first correction filter (42), the filter coefficient of the adaptive FIR filter (41) is changed to the adaptive algorithm execution circuit (4
It is updated sequentially by 4). By updating this filter coefficient, the inversion sound signal is generated by the adaptive FIR filter (41) with respect to the noise (11) in the duct (20) with high accuracy over time, and the noise in the duct (20) ( 11) is best reduced.

At the time of this silencing operation, the adding circuit (45)
Is input with the inverted sound signal output from the adaptive FIR filter (41) and subtracts the inverted sound signal from the noise signal,
Howling is prevented from occurring due to the inversion sound radiated from the speaker (31) propagating to the detection microphone (30).

Next, the prediction operation of the predictor (50) will be described based on the control flow of FIG. First, in step ST1, the data processing circuit (51) acquires an input signal, that is, the noise signal from the detection microphone (30) is acquired by the data processing circuit (51) via the addition circuit (45).

Subsequently, in step ST2, the data processing circuit (51) outputs the predicted noise signal, and the adaptive FIR filter (41) or the like receives the predicted noise signal, while the predicted noise signal is delayed by the delay circuit. Delayed by one sampling time by (54). Then, the process proceeds to step ST4 to compare the delayed predicted noise signal with the noise signal (53).
Compares and outputs a miscalculation signal to the coefficient calculation circuit (52).

Then, the process proceeds to step ST5, and the coefficient calculation circuit (52) calculates the weighting coefficient of the neural network based on the above equations (5) to (16) and outputs the coefficient signal to the data processing circuit (51). To do. And step ST6
And wait until the next sampling, then follow the steps above
Returning to ST1, the above-described operation is repeated, and the data processing circuit (51) outputs the predicted noise signal of one sampling destination, and the coefficient calculation circuit (52) updates the weighting coefficient of the neural network, while the predicted noise signal is calculated. An inverted sound signal is generated based on

-Muffling Result- Next, the result of the muffling experiment of the active muffling apparatus (10) including the above-described predictor (50) will be described.

First, this experiment was conducted on the active silencer (10) shown in FIG. This active silencer (1
0) is an amplifier (33, 34, 35) and an analog filter (3) between the detection microphone (30), the speaker (31) and the monitor microphone (32) and the controller (40).
6, 37, 38) and a delay buffer (46) on the input side of the adder circuit (45).

The delay buffer (46) delays the noise signal of the detection microphone (30) instead of shortening the distance between the detection microphone (30) and the speaker (31). The noise signal is output after the predetermined number of samplings. In this experiment, the sampling frequency was 6
1 kHz for the predictor (50) neural network
There are three layers with two intermediate layers. Further, the first correction filter (42) is omitted.

FIG. 16 shows the attenuation level dB (A) when the delay amount of the delay buffer (46) is 0 sample, 1 sample and 2 samples. In FIG. 16, ▲ represents a conventional configuration in which the predictor (50) is not provided, and Δ, □, and ◯ are cases in which the predictor (50) of the present invention is provided, △ indicates the number of units in the input layer of the neural network is 3, the number of units in the intermediate layer is 3, and the number of units in the output layer is 1. □ is the number of units in the input layer of the neural network, 6 and the intermediate layer. The number of units in is 3 and the number of units in the output layer is 1.
The number of units in the input layer of the neural network is 8, the number of units in the intermediate layer is 3, and the number of units in the output layer is 1.

As is clear from FIG. 16, the ones provided with the predictor (50) (Δ, □, ◯) have a lower attenuation level than those not provided with the predictor (50) (▲). Large and surely attenuated, and also with the predictor (50) installed (△, □, ○), the number of units in the input layer of the neural network is set to 6 or 8 (□, ○) Are almost equal in the attenuation level, and the attenuation level is superior to the one in which the number of units in the input layer of the neural network is set to 3 (Δ).

Further, a predictor (50) having a delay amount of 0 sample
The ones without ▲ (▲) and the ones with a predictor (50) with a delay of 1 sample (△, □, ○) almost correspond to the attenuation level, and the delay is 1 sample. (A) without the predictor (50) of (2) and those with the predictor (50) having a delay amount of 2 samples (□, ○) substantially correspond in attenuation level. That is, since the predictor (50) predicts the noise signal ahead of one sampling time and outputs the predicted noise signal, it can be seen that the predictor (50) reliably predicts the noise signal. It

FIGS. 17 and 18 show changes in the filter coefficient of the adaptive FIR filter (41) due to the difference in delay time. First, the horizontal axes of FIGS. 17 and 18 represent the number of taps of the adaptive FIR filter (41), and the vertical axis represents the filter coefficient. 17 (a), 17 (b), and 18 (c) show the case where the delay amount of the delay buffer (46) is 0 sample, 1 sample, and 2 samples, and FIG. FIG. 18 shows a case where the predictor (50) of the present invention is provided in the conventional case where (50) is not provided. The predictor (50) of FIG. , The number of units in the intermediate layer is 3, and the number of units in the output layer is 1.

In FIG. 17 and FIG. 18, the tap position of the first peak point P shows the time margin until the output, and when comparing FIG. 17 and FIG. I understand. That is, FIG. 17A and FIG.
Since (b) corresponds and FIG. 17 (b) corresponds to FIG. 18 (c), it is possible that the predictor (50) reliably predicts the noise signal ahead by one sampling time. I understand.

FIG. 19 is a comparison between the actual noise signal and the predicted noise signal predicted by the predictor (50). The solid line indicates the noise signal, and the broken line indicates the predicted noise signal. I know that it corresponds.

-Effects of the predictor (50) -As described above, according to the present embodiment, since the noise signal is predicted, the distance between the detection microphone (30) and the speaker (31) is reduced. Since the time corresponding to the estimated time can be shortened, the overall size of the device can be made compact. As a result, the degree of freedom of installation can be improved and the range of application can be expanded.

Further, since the number of units of the input layer in the neural network is set based on the interphase dimension for determining the chaos of the time series data, it becomes easy to express the deterministic relationship in the neural network, and the prediction The accuracy of the noise signal can be improved. As a result, the attenuation level of the noise (11) can be improved.

Since the predictor (50) predicts the noise (11) one sampling ahead, the data processing circuit (51) for updating the weighting coefficient can be omitted. The configuration of the container (50) can be simplified.

[0116]

Second Embodiment A second embodiment of the present invention shows another embodiment of the predictor (50) as shown in FIG. 4, in which time-series data at N sampling destinations are predicted. Is.

That is, the noise signal is input to the prediction data processing circuit (51) corresponding to the data processing circuit (51) of the previous embodiment, while learning the weighting coefficient,
A buffer (55) for delaying the noise signal by N samplings is provided, and the delayed noise signal is input to the weight update data processing circuit (56) via the buffer (55).

Further, the weight update data processing circuit (5
A comparison circuit (53) is provided which compares the predicted noise signal predicted by 6), that is, the predicted noise signal at the N sampling destination and the current noise signal, and outputs an error signal. Then, the coefficient calculation circuit (52) calculates the weight coefficient based on the error signal of the comparison circuit (53), and the weight update data processing circuit (56) updates the weight coefficient.

The weight update data processing circuit (5
Data processing circuit for prediction (51)
Will be copied and the original predicted noise signal will be output.

The specific prediction operation will be described with reference to FIG. 6. First, in step ST11, the prediction data processing circuit (51) acquires an input signal which is a noise signal, and then the process proceeds to step ST12 to predict. A data processing circuit (51) outputs a predicted noise signal. On the other hand, step ST1
In step 3, after the noise signal is input to the buffer (55), the process proceeds to step ST14, and the delay update data processing circuit (56) acquires delay data which is the delayed noise signal.

Then, the process proceeds to step ST15, and the weight update data processing circuit (56) outputs the delayed predicted noise signal at N sampling destinations, that is, predicts the current noise signal and outputs the predicted noise signal. To do. Then, step
After moving to ST16, the comparison circuit (53) compares the predicted noise signal with the current noise signal to output an error signal, and then moves to step ST17 where the coefficient calculating circuit (52) calculates a weighting coefficient and weights it. The update data processing circuit (56) updates the weighting coefficient, and the prediction data processing circuit (51) copies the weighting coefficient, and the data processing circuit (51) outputs the original predicted noise signal.

After that, the process proceeds to step ST18, waits until the next sampling, and then returns to step ST11 to repeat the above-mentioned operation to make the prediction data processing circuit (51) N.
The predicted noise signal at the sampling destination is output, and the inverted sound signal is generated based on the predicted noise signal.

Therefore, in the second embodiment as well, the predicted noise signal is output as in the previous embodiment, so that the entire apparatus can be made compact. However, two data processing circuits (51, 56) are required, and the amount of calculation increases. Therefore, it is preferable to predict the noise (11) one sampling ahead, as in the first embodiment shown in FIG.

[0124]

Other Embodiments of the Invention In the above respective embodiments,
Although the muffling of the fan noise (11) in the air conditioning duct (20) has been described, it goes without saying that the present invention is not limited to these noises (11).

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of the present invention.

FIG. 2 is a circuit block diagram of an active silencer.

FIG. 3 is a circuit block diagram of a predictor.

FIG. 4 is a control flow diagram of a predictor.

FIG. 5 is a circuit block diagram of a predictor showing a second embodiment.

FIG. 6 is a control flow diagram of the predictor showing the second embodiment.

FIG. 7 is a frequency characteristic diagram of fan noise.

FIG. 8 is a microphone voltage data diagram of fan noise.

FIG. 9 is a two-dimensional plane analysis diagram of fan noise.

FIG. 10 is an explanatory diagram illustrating a Lyapunov index.

FIG. 11 is an analysis diagram of a Lyapunov index of fan noise.

FIG. 12 is an analysis diagram of a correlation integral of fan noise.

FIG. 13 is a characteristic diagram of an interphase index with respect to a fan noise embedding dimension.

FIG. 14 is an explanatory diagram of a back propagation algorithm.

FIG. 15 is a circuit block diagram of an experimental active silencer.

FIG. 16 is a characteristic diagram of an attenuation level as an experimental result.

FIG. 17 is a characteristic diagram of a filter coefficient as an experimental result according to a conventional example.

FIG. 18 is a characteristic diagram of filter coefficients as an experimental result according to the present invention.

FIG. 19 is a comparison diagram of a noise signal and a predicted noise signal.

FIG. 20 is a circuit block diagram for explaining a conventional active silencer.

FIG. 21 is a circuit block diagram of a conventional adaptive filter.

FIG. 22 is a circuit block diagram for explaining a conventional adaptive filter.

FIG. 23 is another circuit block diagram for explaining the conventional adaptive filter.

FIG. 24 is a circuit block diagram of a conventional active silencer.

[Explanation of symbols]

10 Active silencer 11 Noise 21 Propagation space 30 Detection microphone (noise detection means) 31 Speaker (reverse sound emission means) 32 Monitor microphone (error detection means) 40 Controller 41 Adaptive FIR filter (digital filter) 42 First correction filter 43 Second correction filter 44 Adaptive algorithm execution circuit 50 Predictor (prediction means) 51 Data processing circuit 52 Coefficient calculation means

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification number Internal reference number for FI Technical indication G10K 11/16 9274-5J H03H 17/00 601M H03H 17/00 601 9274-5J 21/00 21 / 00 G10K 11/16 B (72) Inventor Takeshi Nishimura 1304 Kanaoka-machi, Sakai-shi, Osaka Daikin Industry Co., Ltd. Kanaoka factory (72) Inventor Hiroyuki Ito 1304, Kanaoka-machi, Sakai-shi, Osaka Sakai Daikin Industries, Ltd. Kanaoka Factory

Claims (4)

[Claims] 1. A noise detection means (30) for detecting a noise (11) propagating through a propagation space (21) and outputting a noise signal, and a noise detection means (30) for receiving a noise signal output by the noise detection means (30) The noise (11) that is ahead of the noise signal in time is predicted, and the prediction means (50) that outputs the predicted noise signal and the predicted noise signal that is output by the prediction means (50) have the same amplitude in opposite phase. A digital filter (41) for generating and outputting a reversal sound signal, and a reversal sound radiating means (31) for receiving a reversal signal from the digital filter (41) and radiating a reversal sound to the propagation space (21), Error detecting means (32) for detecting an error between the reversing sound emitted by the reversing sound emitting means (31) and the noise (11) in the propagation space (21) and outputting an error signal, and the error detecting means (32) The error signal is reduced by receiving the error signal output by and the prediction noise signal output by the prediction means (50). And an adaptive algorithm execution circuit (44) for updating the filter coefficient of the digital filter (41). 2. The active silencer according to claim 1, wherein the predicting means (50) predicts the noise (11) at one sampling destination and outputs a predicted noise signal. . 3. The active noise suppressor according to claim 2, wherein the prediction unit (50) receives the noise signal and outputs the predicted noise signal based on a neural network, the predicted noise signal and the data processing circuit (51). And a coefficient calculation circuit (52) for calculating the weighting coefficient of each unit in the neural network based on the difference between the predicted noise signal and the actual noise signal and outputting the coefficient signal to the data processing circuit (51). An active muffling device characterized in that 4. The active silencer according to claim 3, wherein the data processing circuit (51) determines the number of units in the input layer based on the interphase dimension for determining the chaoticity of the noise signal. Active silencer.