JP2002023766A - Method and device for reducing cyclic noise - Google Patents

Abstract

PROBLEM TO BE SOLVED: To realize sound pressure reduction at the target position of the cyclic noise of a fan or the like by adopting a learning control as a control algorithm substitutable for a conventional filtered-X (one input/one output system) or a filtered-ε (multi-input/multi-output system). SOLUTION: The template of cyclic components set based on a reaching sound at the target position for noise reduction from a noise source which generates a noise including a cyclic sound is stored in a storage device. When controlling the noise reduction, a sound detected by a sound receiving device placed at the target position of the noise reduction is compared with the template, and an arithmetic operation using learning control is performed so that the difference between them can be reduced, and a sounding device is controlled based on the arithmetic result. In this case, up to the primary and secondary components of the cyclic components of the noise source are enough to be included in the template.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for reducing noise mainly composed of a periodic component, which is emitted from a fan or a pump of an air conditioner or the like.

[0002]

2. Description of the Related Art Fan noise generally consists of a wideband frequency component and a periodic component. For certain types of fans, the periodic component is dominant, and this component dominates the overall level. For example, the cross-flow fan noise often used in air conditioners has a remarkable periodic component,
Dominates overall fan noise. Therefore, in this type of fan, reducing the periodic noise component immediately leads to a reduction in the overall noise level. For example, the periodic sound of a cross flow fan is caused by fluctuations in the pressure distribution of the blades induced by the interaction when the blades of the rotor pass through the stationary part of the fan. Since such an interaction is inseparable from the pressure increase mechanism of the fan, eliminating pressure fluctuations leads to a reduction in pressure increase. That is, with this type of fan, it is difficult to reduce the generated fan noise itself while maintaining the pressure rise, which is the basic performance specification. Therefore, various noise reduction methods have been considered for reducing noise at a target position where the generated fan noise reaches.

[0003] As one of such silencing methods, feedforward active control has already been developed and many successful examples have been reported (eg, PA Nelson, SJ Ellio).
t: "Active Control of Sound", Academic Press, (199
2), Masaharu Nishimura: "Duct silencing control by ANC, Journal of the Japan Society of Precision Engineering, Vol. 64 (1998) No. 5, pp. 669-673).

Another method is active control (active control). For active control, filtered-X has been applied to a one-input one-output system and filtered-ε has been applied to a multiple-input multiple-output system as a method of LMS feedforward control (B. Widrow,
E. Walach: "Adaptive Inverse Control", Prentice Ha
ll PTR (1996)).

[0005]

In the present invention, learning control is applied as an excellent control algorithm replacing these.

The learning control was originally developed by Kawamura / Miyazaki / Arimoto for use in controlling the operation of a robot (Sadao Kawamura, Fumio Miyazaki, Taku Arimoto: "Learning control method for dynamic systems (betterment) Proposal of process) ", Transactions of the Society of Instrument and Control Engineers, Vol.22 (1986), No.1, pp.56-62). For example, assume a task of drawing characters with a robot arm. In this case, a character template is provided in advance. By repeating the improvement by trial and error, it is possible to acquire a time series of a robot operation that draws a character almost infinitely close to a template in a finite number of times. The present invention introduces this learning control algorithm into the active control of the periodic fan noise.

[0007]

That is, the method for reducing periodic noise according to the present invention comprises the steps of: a) previously determining the arrival sound at a noise reduction target position from a noise source generating noise including periodic sound; Based on a template of a periodic component, b) at the time of noise reduction control, a sound detected at the target position is compared with the template, and learning control is used so as to reduce a difference between the two. Performing a calculation, and c) controlling the sounding device based on the calculation result.

[0008] The apparatus for performing such noise reduction control includes: a) a periodicity set in advance based on a sound that arrives at a noise reduction target position from a noise source that generates noise including periodic sound. A storage device for storing a component template (sample); b) a sound receiving device provided at the noise reduction target position; c) a sound generating device; d) a sound receiving device which is detected by the sound receiving device during noise reduction control. And a control device that performs a calculation using learning control so as to reduce the difference between the sound and the template, and controls the sounding device based on the calculation result. be able to.

In these control methods and apparatuses, it is desirable to employ up to the first and second order components of the periodic component of the noise source in the template. Of course, higher order components may be employed, but the calculation becomes complicated and the response speed is reduced, but it does not lead to a large noise reduction effect.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [Principle of Learning Control]
Temporal repetition phenomena realized at [0, T]
Rate is y_dLet it be given by (t). This is the control input
Time series u_kAchieve the template phenomenon by (t)
Try to. In learning control, y_dEnter goals that can achieve (t)
Force u _dFind (t) by successive iterations. like this
Learning course at the kth time u_{k + 1}= F (u_k, Δy_k)
In other words, the output error Δy in the repetition process_k(t) = y
_k(t) -y_dWhat is necessary is to make (t) converge to 0. this
The learning process is expressed by the following equation (1).

(Equation 1) Φ may be a positive definite matrix composed of constants. Such a learning process is shown in FIG. Function in FIG. 1 F (u _k, Δy
It can be understood that Expression (1) including the matrix Φ has been given as a specific example of _k ). In the following, the conditions under which this iterative learning converges will be examined.

Now, the error amount of u _k (t) is expressed as Δu _k (t) ≡Δu
It is defined as _{k (t) -u d (t} ). At this time, the following relationship is obtained from Expression (1).

(Equation 2) By integrating the inner product of both sides of the above equation (2) and taking the respective norms, the following relationship is obtained.

(Equation 3) Here, the norms ‖ and ‖ are defined as follows.

(Equation 4) Here, it is assumed that the following inequality holds for the third term on the right side of Expression (3).

(Equation 5) Where β is a positive constant and V _k (t) is a non-negative function. If such a relationship holds, then the equation
The following relationship is obtained from (4).

(Equation 6)Where V_k≡V (kT). The inequality above is
Means That is, as the number of trials k increases, β>
If 0, ‖Δu_k‖^Two _Φ-1+ V_kIs ‖Δy_k‖_ΦIs nothing
It decreases monotonically unless it goes away. From this, k → ∞
‖Δy_k‖ _Φ→ O can be proved
You.

When β is positive and the above proof holds, it means that “passivity” holds in the dynamic system to be controlled. Since the passivity is also established in the acoustic system according to the present invention, the convergence of the learning control can be guaranteed if the matrix Φ is positively definite.

[Learning control of fan noise] As described above, learning control is effective only for periodic sounds. Therefore, only the periodic component is extracted in real time from the fan noise including the wide frequency band component, and the learning control is applied to the component. The procedure is shown below.

(1) Extracting the Fourier component of the rotating sound The Fourier component fluctuates in both frequency and Fourier coefficient in an actual phenomenon. Here, the primary and secondary rotation sounds
The components up to the next component are to be controlled, and the error signal p (t) detected from the microphone to be controlled is expressed as follows.

(Equation 7) Here, w (t) is a component of random fluctuation. Strictly speaking, there can be third- or higher-order components of the rotating sound, but as shown later, these can hardly be distinguished in this experiment,
This modeling is considered valid. The fundamental frequency f ₀ or the fundamental period T = 1 / f ₀ also fluctuates, but in practice the change is very gradual, so that it is treated as a constant value during control. In order to extract only the periodic component from this signal in real time, the following finite-time Fourier coefficient is calculated.

(Equation 8) Here, the above equation (8) shows that the Fourier coefficients were calculated in finite time mT. According to the results of a simple numerical experiment, m = 9
It was found that the Fourier coefficients could be calculated with sufficient accuracy for practical use.

[0015] (2) Equation (1) to determine the coefficient matrix Φ control input and an output error vector △ y _k (t) and the vector u _{k + 1} (t), associate u _k. Here, the coefficient matrix Φ may be a positive definite square constant matrix or a constant matrix having all positive singular values. From the relationship of equation (1), it can be seen that Φ has the same meaning as the inverse matrix of the transfer function from the control input to the sensor output.

In this experiment, a filtered ε control experiment was performed as one of the adaptive feedforward controls in order to compare the learning control with the learning control. It was determined. In this case, it has been confirmed that the inverse matrix has a positive definite or positive singular value.

[0017]

For example, in the case of fan noise, pressure fluctuation as sound is repeated at high speed. The active control system only has to use the original fan sound as a template and realize the pressure fluctuation of the secondary sound in the opposite phase of the template in finite learning. In this case, the signal of the error microphone of the conventional feedforward control system is used as it is. However, no reference signal is required here. Also,
In conventional feedforward control, the impulse response that minimizes the error in the least squares sense is identified, and the FIR
Regardless of the model or IIR model, it took a long time and a large program to converge the identification. On the other hand, in the learning control, the conditions for the convergence of the learning for realizing the template are specified, and no instability is induced to speed up the convergence.

[0018]

FIG. 2 shows a model of a wall-mounted air conditioner 11 having a built-in cross flow fan, and an active control system for noise reduction according to the present invention applied thereto. The fan is driven by an AC motor, and the rotation position of the rotor is detected by a rotation sensor (RC) 18 as a signal synchronized with the position of the blade. The air conditioner 11 was mounted close to the laboratory wall so that the noise source of the fan was considered embedded in the baffle rigid wall. At four locations above and below the air conditioner 11, control speakers 12 as secondary sound sources are attached. One or two condenser microphones (M
IC) 13 is located at a radius of about 1.2 m in front of the fan. The detection signal of the MIC 13 is supplied to a microphone amplifier (MIC)
AMP) 14 and a digital signal processor (D
SP) 15. Note that a personal computer (PC) 16 is used as a monitor device of the digital signal processor (DSP) 15.

Digital signal processor (DSP) 1
5 processes a detection signal by the MIC 13 according to a program created in advance based on the learning control, and creates a control signal. The transmitted control signal is transmitted to two amplifiers (AM
The signal is sent to the control speaker 12 via P) 17, where the sound is produced.

[Regarding Generation of Fan Sound Period Component] The generation of the periodic sound of the cross flow fan is caused by the pressure fluctuation experienced by each blade during one rotation or as a result of integrating the pressure fluctuation over one blade. Fluid force fluctuations radiate the sound field of the rotating dipole, expressed as the sum over all the blades. Assuming that the rotor has z identical blades,
Since there is a phase difference of 2π / z between one period, the fundamental frequency of the rotation sound observed in the far sound field is f ₀ = zn / 60 (n: rotation speed
[rpm]). FIG. 3 shows an example of the actually measured value of the noise spectrum. In the case of this fan, z = 36, n = 933 [rp
m], and f ₀ = 560 [Hz]. Here, only the fundamental frequency component and its double component are prominent.

[Result of Learning Control by PC] A program using the algorithm of the learning control according to the present invention was mounted on the air conditioner model 11 of FIG. The control secondary sound source is composed of four speakers 12, and as shown by the connection to the AMP 17,
The same input signal is given to each of the speakers and 2
Considered as an input system.

Two output error microphones 13 were also used, and one was placed at the front and another at a vertical angle of 40 degrees with a radius of 1.2 m in a vertical section passing through the center of the fan. In order to identify the frequency of the fan rotation sound, the rotation speed of the rotation shaft was detected. Here, the fundamental frequency of the rotating sound is f ₀ = 560
At first, it was adjusted to [Hz]. The sample rate of the computer program at the time of control was set to 10 [kHz].

In this case, FIG. 3 shows the frequency analysis result of the fan sound after the learning control is converged by the error microphone 13 in front of the fan (φ = 0 [deg]). As is clear from the comparison before the control (dotted line) and after the control (solid line), a decrease of about 12 [dB] was obtained in the primary component of the rotating sound. In addition, a reduction of about 2 to 3 [dB] was obtained for the second-order component.

In this case, the convergence of the rotation sound by the learning control
FIG. 4 shows the change of the coefficient of the first-order component to show the transition.
In FIG. 4, the horizontal axis represents the number of sample steps, and the vertical axis represents the Fourier relation.
Number c ₁Are shown respectively. From this, 1400 steps (0.1
4sec), which is almost converged,
Which adaptive feedforward control compared to the adaptive speed
This is a much faster response.

The purpose of noise reduction is not achieved simply by reducing the sound pressure level at the error microphone location. The ultimate goal is to reduce sound pressure in a hemisphere with a radius of 1.2 m and to eliminate sound waves traveling outside the hemisphere. In order to investigate this effect, the directivity of sound pressure at a radius of 1.2 m in a vertical plane including the center of the fan is shown in FIG.
FIG. 6 shows the directivity of sound pressure at a radius of 1.2 m in a horizontal plane including the center of the fan. In these figures,
The directivity after control represented by a solid line is shown in comparison with that before control represented by a dotted line. Thus, in general, the control effect is more remarkable in the directivity in the vertical plane than in the horizontal plane. This seems to be due to the effect of arranging the two microphones 13 at two places in the vertical plane. Further, FIG. 7 shows directivity in the same vertical plane when filtered-ε is applied as a control algorithm under the same microphone arrangement. Than this,
It can be seen that the effect of learning control is slightly larger than the effect of filtered-ε.

Finally, learning control was applied to the active control of the rotation noise generated in the cross flow fan. For the primary and secondary components, a reduction effect of a maximum of 15 [dB] was obtained at the sound pressure level. In addition, a reduction effect was observed on almost the entire surface of the hemisphere having a radius of 1.2 m from the center of the fan.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a learning process performed by learning control according to the present invention.

FIG. 2 is a schematic configuration diagram of a wall-mounted air conditioner and its control mechanism used in a performance experiment of the noise reduction control system according to the present invention.

FIG. 3 is a frequency spectrum of a measured fan sound value before and after convergence of the noise reduction learning control of the air conditioner used in the experiment.

FIG. 4 is a graph showing a change in a coefficient of a primary component of a Fourier component in learning control.

FIG. 5 is a graph showing the directivity of sound pressure at a radius of 1.2 m in a vertical plane including the center of the fan.

FIG. 6 is a graph showing the directivity of sound pressure at a radius of 1.2 m in a horizontal plane including the center of the fan.

FIG. 7 is a graph showing directivity in the same vertical plane as in FIG. 5 when filtered-ε is applied as a control algorithm under the microphone arrangement of FIG. 2;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 ... Air-conditioner model 12 ... Control speaker 13 ... Error microphone (MIC) 14 ... Microphone amplifier (MIC AMP) 15 ... Digital signal processor (DSP) 16 ... Personal computer (PC) 17 ... Speaker amplifier (AMP) 18 ... Rotation sensor (RC)

Claims (5)

[Claims] 1. a) a template of a periodic component is set in advance based on a sound that arrives at a noise reduction target position from a noise source that generates noise including a periodic sound; b) noise reduction control Sometimes, the sound detected at the target position is compared with the template, and a calculation using learning control is performed so as to reduce the difference between the two.c) controlling the sounding device based on the calculation result; A method for reducing periodic noise. 2. The method according to claim 1, wherein the template includes first and second order components of the periodic component of the noise source.
The method for reducing periodic noise described in the above. 3. A storage device for storing in advance a template of a periodic component set based on a sound that arrives at a noise reduction target position from a noise source that generates noise including periodic sound. B) a sound receiving device provided at the noise reduction target position; c) a sound generating device; and d) a sound detected by the sound receiving device and the template during noise reduction control, and a difference between the two is compared. A control device that performs an operation using learning control so as to reduce the noise, and controls the sounding device based on the operation result. 4. The template includes up to first-order and second-order components of a periodic component of a noise source.
The periodic noise reduction device as described in the above. 5. The apparatus according to claim 3, further comprising a plurality of sound receiving devices and a plurality of sound generating devices, wherein the control device performs multi-input multi-output learning control.
Or the periodic noise reduction device according to 4.