JP3150531B2 - Sound quality evaluation device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a sound quality evaluation device which is applied to the evaluation of noise of a general mechanical device that generates noise.
In particular, the present invention relates to a sound quality evaluation device applied to noise evaluation of mass-produced machines such as automobiles, internal combustion engines, air conditioners, and construction machines.

[0002]

2. Description of the Related Art Conventionally, with respect to mechanical noise caused by mass-produced machines such as automobiles and internal combustion engines, the loudness or loudness of the noise is determined based on the magnitude of the volume measured by a noise measuring device. Therefore, the usual noise countermeasures are taken for the purpose of reducing the measured sound volume, and the surrounding environment is protected. Normally, an evaluation criterion in units of “phones” is set for the volume, and the noise measurement measures the number of phones. Therefore, the noise countermeasure aims to reduce the number of phones to be measured to a predetermined level or less prescribed by law.

[0003]

However, in the above-described conventional noise measurement and noise countermeasures, since the main purpose is to reduce the volume, there is no evaluation of sound quality (tone) that cannot be defined by the volume evaluation method. There are no guidelines or equipment to approach, and no proper measurements and measures have been taken. In particular, the sound quality (tone) depends on the sense of each person and is difficult to handle, and therefore, it is difficult to take measures against noise in consideration of the sound quality.

The present invention has been made in view of the above-mentioned circumstances, and it is possible to quantitatively and automatically calculate and evaluate the sound quality evaluation of mechanical noise, thereby improving the accuracy of sound quality evaluation and It is an object of the present invention to provide a sound quality evaluation device capable of speeding up.

[0005]

SUMMARY OF THE INVENTION The sound quality evaluation apparatus according to the present invention measures the noise data of the machine is subject,
A data recording means for converting the noise data into a digital amount and holding the data; and passing only a predetermined critical band with respect to the noise data held by the data recording means, and adjusting a volume of the passed noise data for each critical band. Calculating means for calculating;
According to the mechanical device, a neural network is provided so that the volume of each critical band and the degree of sound quality preference appropriately correspond to each other, and the neural network is calculated by the arithmetic unit.
And evaluating means for evaluating and displaying the degree of sound quality preference based on the volume of each critical band .

The above-mentioned critical band level is a volume passing through a critical band that is deeply related to the human auditory phenomenon, and as shown in FIG. 2, a total of 24 bands in the audible frequency range are representative. Exists as an example.

Furthermore, the neural network in the evaluation means learns in advance the relationship between the critical band level and the sound quality preference for each type of mechanical noise (for example, automobile, internal combustion engine, etc.).

[0008]

The data recording means measures noise data of a mechanical device as a subject. Further, the measured noise data is sampled (discretized) using an analog / digital converter or the like, and stored. The calculating means is a band capable of passing through only the critical band (for example, 24 bands represented in the audible frequency range) which is closely related to the human auditory phenomenon with respect to the noise data held in the data recording means. The critical band level of the passed noise data is calculated using a pass filter.

The evaluation means has been learned in advance so that the critical band level and the degree of sound quality preference appropriately correspond to the mechanical device to be inspected, that is, an appropriate weight and threshold value. (Bias). The evaluation means uses the neural network and evaluates the degree of the sound quality preference of the subject by inputting the critical band level of each critical band obtained by the calculation means. This evaluation result is displayed on a display such as a display. The above-described series of data processing is automated, and only by performing noise measurement, it is possible to evaluate not only the sound volume but also the sound quality which has been difficult in the past.

[0010]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows the configuration of a sound quality evaluation device according to the present invention. As shown in FIG. 1, the noise measuring device 1 measures the sound pressure of mechanical noise radiated from a mechanical device as a subject. The analog / digital converter 2 converts sound pressure data, which is an analog quantity measured by the noise measuring instrument 1, into a digital quantity. The data storage 3 stores the sound pressure data converted by the analog / digital converter 2.
The noise measuring device 1, the analog / digital converter 2, and the data storage device 3 measure the mechanical noise of the subject and record the data.

The digital filter unit 4 allows the sound pressure data stored in the data storage unit 3 to pass each noise component in ₂₄ (F _{1 to} F ₂₄ ) critical bands within the audible frequency range. The 24 critical bands are defined as shown in FIG. 2 and are associated with human auditory phenomena. The critical band level calculator 5 calculates the volume of each band passing through the 24 critical bands. As shown in Figure 1 F ₁ ~
Each sound volume corresponding to the bandwidth of the F ₂₄ is P ₁ to P _24. The neural network preference estimator 6 estimates and evaluates the sound quality preference degree α _e based on each critical band level measured by the critical band level calculator 5. The neural network preference estimator 6 uses a critical band level based on human auditory characteristics as an input and a preference degree indicating sound quality as an output, and the critical band level and the preference appropriately correspond to each other. It consists of weights and thresholds. The display 7 displays the evaluation result by the neural network preference estimator 6 on a display.

The critical band level indicates a volume passing through a critical band that is closely related to a human auditory phenomenon. As described above, each critical band generally has 24 representative examples as shown in FIG. Further, the above-mentioned preference indicates a "preferred manner of sound", and this preference is a statistic, and can be any numerical value according to a definition method.

Next, the basic configuration and operation of the neural network used in the neural network preference estimator 6 will be described with reference to FIG. The neural network is used to approximate a nonlinear function.
A hierarchical network in which basic units (neurons) similar to neurons in the brain as shown in FIG. 3A are connected from the input layer to the output layer so as to form a plurality of layers as shown in FIG. is there. As the input / output function of the unit, a logarithmic sigmoid function (see FIG. 4) capable of converting positive and negative infinite inputs from 0 to 1 can be used.

Here, focusing on the input / output relationship of the units, the i-th unit (i-th unit) of the k-th layer other than the input layer is the same as the j-th unit of the preceding layer (k-1). Weighted inputs are received from all the units including them, and a function output is performed based on their sum. That is, i ^k _i and o ^k _i are the sum and output of the inputs of the i-th unit in the k-th layer, and w _ji is the weight b of the transmission from the j-th unit in the k−1-th layer to the i-th unit in the k-th layer. ^Assuming that ^k _i is the threshold value of the i-th unit in the k-th layer, the following relationship is established.

[0015]

(Equation 1)

When such units are arranged hierarchically and an input pattern is presented to the input layer, conversion is continuously performed in the network, and a conversion result can be obtained in the final output layer. To make this conversion result desirable, there is a learning algorithm for providing a teacher signal to the output layer and updating the weights and thresholds of the units based on the deviation. As a typical example, there is an error back propagation method (Error Back Propagation).

Next, the error back propagation method will be described. Although the input-output relationship of the equation (1) is satisfied at the k layer network m layer, defining an evaluation function r at the output layer and the sum of the square errors output o ^m _n and the teacher signal y _n of the n-th unit Then, it becomes as follows.

[0018]

(Equation 2) Further, the weight correction amount is Δw _ji ,
If ε is a positive small coefficient, the following relationship is obtained.

[0019]

(Equation 3) Here, regarding the right side of the above equation (3)

[0020]

(Equation 4) When k = m, the above equation (2) is used.

[0021]

(Equation 5) Becomes When k = m−1, the first of the (m−1) th layer
Regarding the unit,

[0022]

(Equation 6) Is given by the following.

[0023]

(Equation 7) Therefore, assuming that the (m-1) th layer is the (k + 1) th layer, the recurrence formula for the kth layer is obtained as follows.

[0024]

(Equation 8) Therefore, in the above equation (4),

[0025]

(Equation 9) Then, the amount of correction of the weights is summarized by the above equation (3) as follows.

[0026]

(Equation 10) Also, regarding the threshold, in the above equation (4)

[0027]

[Equation 11] Therefore, the correction amount Δb ^k _i is similarly obtained. In the above equation, the learning signal d ^k _i used for correction is recursively calculated from the output layer (m layer) to the input layer, and convergence calculation is performed until the evaluation function reaches a target value.

Next, an example of the configuration of the neural network preference estimator 6 of the embodiment using the above-described neural network will be described with reference to FIG. As shown in FIG. 5, each volume of the critical band level calculator 5 is used for an input layer, and a logarithmic sigmoid function with a bias and a second
A biased linear function is used for the layers. The first and second layers each receive an input (i ^k _i ) and a next output (o
^k _i ). These relationships are shown in the following equations.

[0029]

(Equation 12)

When the above relationships are arranged hierarchically as shown in FIG. 5, the range from the critical band level of the input to the preference of the output is continuously expressed by a mathematical expression. Therefore,
Neural network shown in FIG. 5, as the relationship of the input and output and a weight Wji and threshold (bias) bi in the above-mentioned relationship is a corresponding, e.g. yield using a technique such as error back propagation analysis described above It can be learned by obtaining by a bundle calculation.

Next, the operation of this embodiment will be described. The sound pressure data of the mechanical noise radiated from the mechanical device as the subject is measured by the noise measuring device 1. The measured sound pressure data is converted into a digital quantity by the analog / digital converter 2 and stored in the data storage 3. In this way, noise measurement and data recording for the mechanical device are performed.

Next, the sound pressure data stored in the data storage 3 is applied to the digital
Only each noise component in the four critical bands is passed.
For the noise component passing through each critical band, the critical band level calculator 5 calculates the volume of each band. Each of the calculated sound volumes is used as an input layer of the neural network preference estimator 6 to calculate a sound quality preference degree α _e . The sound quality preference α _e estimated and evaluated by the neural network preference estimator 6 is displayed on the display 7. With the above operation, the sound quality of the mechanical noise is evaluated.

[0033]

Effect of the Invention] According to the invention, as Shoki, the machine sound quality neural network learning that corresponds has been subjected to, i.e., based on human auditory characteristics to the input extraordinary
The sound volume of each field band is used as a subject by providing a sound quality evaluation neural network composed of a weight and a threshold value that appropriately correspond to weights and thresholds, using a preference degree indicating sound quality superiority or output. If you measure only the noise of mechanical equipment,
It is possible to quantitatively evaluate the sound quality of mechanical noise which has conventionally depended only on human sensation, and to automatically calculate and evaluate. Accordingly, it is possible to improve the accuracy and speed of sound quality evaluation. Therefore, it can be used for sound quality improvement for environmental conservation.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a sound quality evaluation device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a frequency definition of a typical critical band in hearing applied to the embodiment.

3A and 3B show a basic configuration of a hierarchical network applied to this embodiment, in which FIG. 3A shows a neuron forming the hierarchical network, and FIG. 3B shows a connection form of the neuron.

FIG. 4 is a diagram showing an example of a logarithmic sigmoid function with a threshold used in the hierarchical network.

FIG. 5 is a diagram showing a configuration of a neural network applied to this embodiment for estimating a preference degree of sound quality.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 noise measuring device 2 analog / digital converter 3 data storage device 4 digital filter device 5 critical band level calculator 6 neural network preference estimator 7 display

Continuation of the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G01H 3/00 G01H 17/00 G06F 15/18 560

Claims (1)

(57) [Claims] 1. A noise data of mechanical devices are subject meter <br/> to measurement, and data acquisition means for holding and converts the noise data to digital amount, the noise data held in the data recording means hand, it passes only predetermined critical band, the noise data which has passed through extraordinary
Calculating means for calculating a sound volume for each of the boundary bands, and a neural network which is provided so that the sound volume for each of the critical bands and the degree of sound quality preference appropriately correspond to the mechanical device. Calculated by means
An evaluation means for evaluating and displaying a degree of sound quality preference based on the volume of each critical band .