JP6910641B2 - Small speaker design support device and speaker design support method - Google Patents

Description

The present disclosure relates to a small speaker design support device and a speaker design support method, and more particularly to a technique for supporting the design of a speaker used in a small device such as a smartphone.

In recent years, the demand for small devices such as smartphones has expanded, and along with this, the demand for speakers built into small devices has also increased, and the speakers have become smaller and more complex.

Conventionally, a so-called acoustic equivalent circuit analysis method has been known as a speaker design method. The acoustic equivalent circuit analysis method replaces the mechanical vibration system such as a diaphragm and a damper and the acoustic vibration system such as a cabinet and a horn with an equivalent electric circuit, and uses the knowledge of the electric circuit to use the target acoustic equipment (speaker). Etc.) is a method for analyzing the acoustic characteristics. Such an acoustic equivalent circuit analysis method is described in, for example, Non-Patent Documents 1 and 2.

Edited by Takeo Yamamoto, "Speaker System (1)", Radio Technology Co., Ltd., July 15, 1977, p. 35-44 "Acoustic Engineering (21st Century Oriented Electronics, Communication, Information Curriculum Series)" by Atsuro Mitsuida, Shokodo, March 1987, p. 32-39 Edited by Masahiro Toyoda, edited by Acoustical Society of Japan, "The World of Sounds Seen by the FDTD Method (Acoustic Science Series 14), Corona Publishing Co., Ltd., December 16, 2015, p.1-46"

In a small speaker that is built into a small device such as a smartphone, a diaphragm that serves as a sound source is provided inside the housing, and the dimensions and positions of the sound holes provided in the housing, as well as the diaphragm and sound hole. The dimensions of the air chamber provided between and the air chamber are important design parameters. However, since the above-mentioned acoustic equivalent circuit analysis method cannot consider the size and position of the sound hole, the above-mentioned analysis method does not need to consider the size and position of the sound hole in particular. However, it cannot be applied to the design of a small speaker in which the size and position of the sound hole are important design parameters. For this reason, in the design of a small speaker, a cut-and-try design is performed by an engineer who has sufficient experience, and reduction of time and cost required for the design is an issue.

The present disclosure has been made to solve such a problem, and an object of the present invention is to provide a design support device capable of reducing the time and cost required for designing a small speaker.

The design support device of the present disclosure is a design support device for a small speaker, and includes an input device and an arithmetic unit. The input device is a device for the user to input acoustic characteristic data indicating the acoustic characteristics of the small speaker. The arithmetic unit now calculates the design parameters from the acoustic characteristic data input by the input device, using a trained model to make the computer function to output the design parameters of the small speaker based on the acoustic characteristic data. It is composed. Here, the trained model is generated by computer simulation using the acoustic FDTD (Finite-Difference Time-Domain) method, and includes a plurality of design parameters and acoustic characteristic data of a small speaker having the design parameters. It is composed of a deep neural network (hereinafter referred to as "DNN (Deep Neural Network)") trained to calculate design parameters from acoustic characteristic data using training data.

Preferably, the small speaker is configured to include a diaphragm serving as a sound source and a space inside the housing formed inside the housing of the small speaker. The size of the divided grid in the computer simulation using the acoustic FDTD method is smaller in the order of the diaphragm, the space inside the housing, and the acoustic propagation space outside the housing.

Preferably, the design parameters included in each learning data are randomly generated within a predetermined range.

Preferably, the acoustic characteristic data includes data indicating the frequency characteristic of sound pressure at a predetermined analysis point.

Preferably, as the acoustic characteristic data used as a plurality of learning data, either data showing the frequency characteristic of the absolute value of the sound pressure or data showing the frequency characteristic of the sound pressure level is selected according to the request of the user. NS.

Preferably, the acoustic characteristic data further includes data indicating the frequency characteristic of the phase at the analysis point.

Preferably, the design parameters include the dimensions and location of the sound holes formed in the small speaker as well as the dimensions of the air chamber.

Further, the design support method of the present disclosure is a speaker design support method, which includes a step of generating a trained model for operating a computer to output a speaker design parameter based on acoustic characteristic data, and acoustics. It includes a step of inputting characteristic data by the user and a step of calculating design parameters from the input acoustic characteristic data using the trained model. The step of generating the trained model includes a step of generating a plurality of training data including a design parameter and acoustic characteristic data of a speaker having the design parameter, respectively, by computer simulation using the acoustic FDTD method, and a plurality of steps of generating the trained model. It includes a step of generating a trained model by training the DNN so as to calculate the design parameters from the acoustic characteristic data using the training data.

In the small speaker design support device of the present disclosure, input is performed by an input device using a trained model for operating the computer to output the design parameters of the small speaker based on the acoustic characteristic data of the small speaker. Design parameters are calculated from the acoustic characteristic data. The trained model is composed of DNNs trained to calculate design parameters from acoustic characteristic data. This makes it possible to design a small speaker without relying on a cut-and-try design by an engineer.

Here, a plurality of learning data used for DNN learning, including design parameters and acoustic characteristic data of a small speaker having the design parameters, are generated by computer simulation using the acoustic FDTD method. As a result, it is possible to prepare a large number of learning data used for learning DNN without actually creating various acoustic structures to prepare a plurality of learning data and performing actual measurement.

As described above, according to the design support device for the small speaker of the present disclosure, various acoustic structures are actually created in order to prepare the learning data of DNN without relying on the cut and try design by the engineer. Since it is not necessary to actually measure the speaker, the time and cost required for designing the small speaker can be significantly reduced.

It is a block diagram which shows the structure of the design support apparatus of a small speaker according to the embodiment of this disclosure. FIG. 5 is a cross-sectional view of a small speaker designed by using the design support device shown in FIG. FIG. 5 is a plan view of the small speaker shown in FIG. 2 as viewed from the x direction. It is a figure which showed the design parameter of the small speaker shown in FIGS. 2 and 3 and an example of the range of each parameter. It is a figure which showed the frequency characteristic of the sound pressure (Pa) at the analysis point P0 shown in FIG. It is a figure which showed the frequency characteristic of the phase of the sound wave at the analysis point P0 shown in FIG. It is an whole block diagram of the learning system which generates the trained model used in the arithmetic unit shown in FIG. It is a block diagram of the generation system which generates the learning data used for the learning system shown in FIG. 7. It is an outline drawing when a small speaker is viewed from the x direction. It is an outline drawing when a small speaker is viewed from the x direction. It is a figure explaining an example of the concept of grid setting by a condition setting part. It is a block diagram of the DNN used for the trained model generated by the DNN learning unit. It is a figure which showed the input / output of DNN in this embodiment. It is a flowchart explaining the generation procedure of the trained model used in the arithmetic unit shown in FIG. It is a flowchart explaining the procedure of the process executed in the arithmetic unit shown in FIG. It is a figure which showed the frequency characteristic of the absolute value (Pa) of a sound pressure. It is a figure which showed the frequency characteristic of the sound pressure level (dB).

Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. The same or corresponding parts in the drawings are designated by the same reference numerals, and the description thereof will not be repeated.

<Overall configuration of design support device>
FIG. 1 is a block diagram showing a configuration of a design support device for a small speaker according to an embodiment of the present disclosure. With reference to FIG. 1, the design support device 1 uses a computer to design a small speaker built in a small device represented by a smartphone. The design support device 1 includes an input device 10, an arithmetic unit 20, and a display device 30.

The input device 10 is a device for a user of the design support device 1 who is a designer of the small speaker to input data (acoustic characteristic data) indicating the acoustic characteristics of the small speaker to be designed. The acoustic characteristics of the small speaker can be represented by the sound pressure at a predetermined analysis point, the frequency characteristics of the phase, and the like. In this embodiment, the user can input desired sound pressure frequency characteristics and phase frequency characteristics from the input device 10 for the small speaker to be designed. The acoustic characteristic data will be described in detail later.

The arithmetic unit 20 includes a CPU (Central Processing Unit), a ROM (Read Only Memory) for storing a processing program and a learned model described later, a RAM (Random Access Memory) for temporarily storing data, and the like. .. The trained model will be explained in detail later, including how to generate it, but this trained model should output the design parameters (sound hole dimensions, position, etc.) of the small speaker based on the acoustic characteristic data. It is a model for operating a computer (computing device 20), and is composed of a DNN in which a causal relationship between acoustic characteristic data and design parameters has been learned. Using this trained model, the arithmetic unit 20 calculates the design parameters of the small speaker from the acoustic characteristic data received from the input device 10.

The display device 30 is a device for displaying the design parameters of the small speaker calculated by the arithmetic unit 20 to the user. A user who has input desired acoustic characteristic data from the input device 10 can confirm the design parameters calculated by the arithmetic unit 20 or the structure of a small speaker having such design parameters on the display device 30. The display device 30 may be integrally configured with the input device 10 by, for example, a touch panel display or the like.

<Explanation of design parameters and acoustic characteristic data>
First, the configuration and design parameters of the small speaker designed by using the design support device 1 shown in FIG. 1 and the acoustic characteristic data showing the acoustic characteristics of such a small speaker will be described.

2 and 3 are block diagrams of a small speaker designed by using the design support device 1 shown in FIG. FIG. 2 shows a cross-sectional view of the small speaker, and FIG. 3 is a plan view of the small speaker shown in FIG. 2 as viewed from the x direction. The configuration of the small speaker shown in FIGS. 2 and 3 is a schematic representation of the small speaker focusing on important parameters (dimensions and positions of sound holes, etc.) in the design of the small speaker.

The small speaker 200 designed by using the design support device 1 with reference to FIGS. 2 and 3 includes a housing 210, a diaphragm 220, a sound hole 230, and an air chamber 240. The acoustic propagation space 250 is a space in which sound waves output to the outside of the housing 210 propagate through the sound holes 230. The boundary 260 is an absorbing boundary introduced to make the analysis target region finite in a computer simulation by the acoustic FDTD method described later, and is, for example, a perfectly matched layer (PML).

The diaphragm 220 is a sound source of the small speaker 200. That is, sound waves are generated by the vibration of the diaphragm 220. Although not particularly shown, a voice coil may be used as the drive source of the diaphragm 220 (piezoelectric conversion method), or a piezoelectric element that expands and contracts according to the applied voltage may be used (piezoelectric conversion method). ..

The sound hole 230 is formed in the housing 210, and the sound wave generated by the diaphragm 220 is output to the sound propagation space 250 outside the housing 210 through the sound hole 230. The air chamber 240 is provided between the space where the diaphragm 220 is provided and the sound hole 230. In this embodiment, only one air chamber 240 is provided, but a plurality of air chambers may be provided in the x direction.

LHx (FIG. 2) shows the length of the sound hole 230 in the x direction. Further, LHy (FIGS. 2 and 3) indicates the length of the sound hole 230 in the y direction, and LHz (FIG. 3) indicates the length of the sound hole 230 in the z direction. LCx (FIG. 2) shows the length of the air chamber 240 in the x direction. Hymin and Hzmin indicate the positions of the sound holes 230 in the housing 210. For example, if Hymin = 0 and Hzmin = 0, the position of the sound hole 230 is the lower left corner of the housing 210 in FIG.

An analysis point P0 is set in the acoustic propagation space 250 (FIG. 2). The analysis point P0 is a point for measuring and analyzing the acoustic characteristics of the small speaker 200. The acoustic characteristics of the small speaker 200 are represented by the acoustic characteristics at the analysis point P0, and the desired acoustic characteristics are obtained at the analysis point P0. The small speaker 200 is designed so as to be obtained. In this embodiment, the analysis point P0 is set at a position separated by L0 in the x direction from the center of the sound hole 230 on the end face of the housing 210.

FIG. 4 is a diagram showing an example of the design parameters of the small speaker 200 shown in FIGS. 2 and 3 and the range of each parameter. With reference to FIGS. 2 and 3 as well as FIG. 4, in this embodiment, the design parameters are LCx, LHx, LHy, LHz, Hymin, Hzmin (all ^10-3 m). That is, LCx indicates the size of the air chamber, LHx, LHy, and LHz indicate the size of the sound hole 230, and Hymin and Hzmin indicate the position of the sound hole 230. In the following, these design parameters may be represented by the design variable vector X = [LCx, LHx, LHy, LHz, Hymin, Hzmin].

The range of each parameter indicates the designable range. In this embodiment, the LCx and LHx ranges are 0.2 to 2.0 (10 ^-3 m), and the LHy and LHz ranges are 0.5 to 3.0 (10 ^-3 m). ), And each range of Hymin and Hzmin is 0.5 to 11.5 (10 ^-3 m).

In the small speaker, the size and position of the sound hole with respect to the air chamber have a particularly large effect on the acoustic characteristics of the small speaker. Therefore, in this embodiment, the width of the air chamber 240 (the length in the y direction and the z direction). Is large and constant, and the width of the air chamber 240 in the x direction and the dimensions and position of the sound hole 230 are set as design parameters.

5 and 6 are diagrams showing an example of acoustic characteristic data of the small speaker 200. FIG. 5 shows the frequency characteristic of the sound pressure (Pa) at the analysis point P0 shown in FIG. 2, and FIG. 6 shows the frequency characteristic of the phase (rad) of the sound wave at the analysis point P0. A Gaussian pulse is given as an input to the diaphragm 220 (FIG. 2) (described later).

With reference to FIGS. 5 and 6, such a profile (frequency waveform) of the frequency characteristic of sound pressure and a profile of the frequency characteristic of the phase are examples of acoustic characteristic data showing the acoustic characteristic of the small speaker 200. Further, the frequency at which the sound pressure peaks (the frequency at which the phase changes suddenly) indicates the resonance frequency of the small speaker 200, and such a resonance frequency is also an example of the acoustic characteristic data of the small speaker 200.

With reference to FIG. 1 again, this design support device 1 is used for designing the small speaker 200 (FIGS. 2 and 3) shown by the design parameters shown in FIG. In the design support device 1, desired acoustic characteristic data (FIGS. 5 and 6) is input from the input device 10 by the user, and a design parameter (FIG. 4) of the small speaker 200 showing such acoustic characteristic data is calculated. Calculated by device 20.

As described above, in the small speaker 200, the dimensions and positions of the sound holes 230 (LHx, LHy, LHz, Hymin, Hzmin), the dimensions of the air chamber 240 (LCx), and the like are important design parameters. However, since the acoustic equivalent circuit analysis method shown in Patent Documents 1 and 2 above cannot consider the size and position of the sound hole, this analysis method particularly considers the size and position of the sound hole. Although it is useful for designing a large or medium-sized speaker that does not need it, it cannot be applied to the design of a small speaker 200 in which the size and position of the sound hole 230 are important design parameters.

Therefore, in the design support device 1 according to this embodiment, an input device is used by using a trained model for operating the computer (arithmetic unit 20) so as to output the design parameters of the small speaker 200 based on the acoustic characteristic data. The design parameters are calculated from the acoustic characteristic data input by 10. The trained model is composed of DNNs trained to calculate design parameters from acoustic characteristic data. As a result, the small speaker 200 can be designed without relying on a cut-and-try design by an engineer.

Here, in order to accurately learn the causal relationship between the acoustic characteristic data by DNN and the design parameters, it is necessary to prepare a large number of learning data. Each learning data is composed of design parameters appropriately set in a range (FIG. 4) in which design parameters can be taken, and acoustic characteristic data of a small speaker 200 having the design parameters. Then, in the design support device 1 according to this embodiment, a plurality of learning data used for learning the DNN are generated by a computer simulation using the acoustic FDTD method. As a result, it is possible to prepare a large number of learning data used for DNN learning without actually creating various acoustic structures to prepare a large number of learning data and performing actual measurement.

Hereinafter, the trained model used in the arithmetic unit 20 of the design support device 1 according to this embodiment will be described.

<Explanation of trained model>
FIG. 7 is an overall block diagram of a learning system that generates a trained model used in the arithmetic unit 20 shown in FIG. With reference to FIG. 7, the learning system 100 includes a database 120, a DNN learning unit 130, and a learning condition setting unit 140.

The database 120 stores learning data 110-1, 110-2, ..., 110-N for generating a trained model in the DNN learning unit 130. Each of the learning data 110-1 to 110-N has the above-mentioned design parameter (design variable vector X) and the acoustic characteristic data (sound pressure frequency characteristic and phase frequency characteristic) of the small speaker 200 having the design parameter. It is a pair of data. The training data 110-1 to 110-N have different design parameters within the range of each parameter shown in FIG.

In order to generate a highly accurate trained model in the DNN learning unit 130, it is important to prepare a large number of training data 110-1 to 110-N for various design parameters within the range that the design parameters can take. .. It takes a lot of time and cost to actually create various acoustic structures and actually measure them in order to prepare a large number of learning data 110-1 to 110-N. Therefore, in this embodiment, a large number of learning data 110-1 to 110-N are generated by computer simulation using the acoustic FDTD method.

FIG. 8 is a configuration diagram of a generation system that generates learning data used in the learning system 100 shown in FIG. 7. With reference to FIG. 8, the generation system 300 includes a design parameter setting unit 310, an acoustic FDTD calculation unit 320, and a condition setting unit 330.

The design parameter setting unit 310 generates a large number of design parameters different from each other within the range shown in FIG. 4 in order to generate a large number of learning data 110-1 to 110-N. For example, the design parameter setting unit 310 generates thousands of different design parameters from each other.

It may take a large load and time to generate a large number of training data from a large number of design parameters. However, in this embodiment, the purpose is to reduce the calculation load of the acoustic FDTD calculation unit 320 that generates the training data. In addition, the symmetry of the design parameters (line symmetry, point symmetry) is taken into consideration, and duplicate operations are avoided. Further, the design parameter setting unit 310 randomly generates a large number of design parameters in order to improve the accuracy of the trained model generated by using the large number of training data.

9 and 10 are external views of the small speaker 200 shown in FIGS. 2 and 3 when viewed from the x direction. Note that FIG. 9 does not show the sound hole 230. With reference to FIG. 9, it is assumed that the outer shape of the small speaker 200 when the small speaker 200 is viewed from the x direction is a square. In this case, considering the symmetry of the shape, it is sufficient to prepare the design parameters included in the region 270 in which the center of the sound hole 230 is shaded, and the center of the sound hole 230 is located. There is no need to generate design parameters contained in other areas.

Therefore, the design parameter setting unit 310 generates a large number of design parameters for the design parameters in which the center of the sound hole 230 is included in the region 270 in consideration of the symmetry of the shape of the small speaker 200. In this way, by considering the symmetry of the design parameters, it is possible to reduce the calculation load and the calculation time of the acoustic FDTD calculation unit 320 that calculates the acoustic characteristic data for each generated design parameter.

Further, referring to FIG. 10, the design parameter setting unit 310 randomly sets the design parameter within the range that the design parameter can take (within the range in which the center of the sound hole 230 is included in the region 270 of FIG. 9 in consideration of symmetry). Generate a large number of design parameters. "Random" means that the design parameters are not changed at a fixed step size, but are arbitrarily changed. By using a large number of training data based on randomly changed design parameters, it is possible to improve the accuracy of the trained model by avoiding the learning from falling into a local solution when generating the trained model by DNN. ..

With reference to FIG. 8 again, the acoustic FDTD calculation unit 320 includes a CPU, a GPU (Graphics Processing Unit), a ROM for storing a processing program, a RAM for temporarily storing data, and the like. Then, the acoustic FDTD calculation unit 320 receives a large number of design parameters generated by the design parameter setting unit 310, and the small speaker 200 has the design parameters for each design parameter by computer simulation using the acoustic FDTD method. Calculate the acoustic characteristic data of the case.

The FDTD method is a method originally used for propagation analysis of electromagnetic waves. Maxwell's equations, which are the dominant equations of electromagnetic waves, are differentiated in time and space, and the electromagnetic field in the analysis space is temporally differentiated by using the leapfrog algorithm. It is a method to obtain the time response of the output point by updating. The acoustic FDTD method is an application of this FDTD method to sound wave propagation analysis, and wave propagation is calculated using sound pressure and particle velocity. This acoustic FDTD method can perform acoustic characteristic analysis in consideration of the shape and positional relationship of the sound hole and the air chamber. The outline of the acoustic FDTD method will be described below (see Patent Document 3 above for details).

In the acoustic FDTD method, the sound pressure and particle velocity in the analysis space are spatially and temporally determined by using a structured grid in which the reference points of sound pressure and particle velocity are arranged in a spatially and temporally staggered manner, which is called a staggered grid. Is discreteized into. Then, using the leap frog algorithm, the spatially and temporally discretized sound pressure and particle velocity are alternately calculated.

In the acoustic FDTD method, the motion and deformation of air particles (small grid-like volumes) in the analysis space are governed by the equations shown below.

Equations (1) to (3) are equations of motion of air particles in the x, y, and z directions, respectively, and equation (4) is a continuity equation relating to sound pressure. ρ indicates the air density, and u, v, and w indicate the particle velocities in the x, y, and z directions, respectively. Further, p indicates sound pressure and κ indicates bulk modulus. The above equations (1) to (4) are collectively referred to as "governing equations".

The sound pressure and particle velocity in the analysis space can be obtained by solving the above governing equation, but in order to realize this on a computer, the acoustic FDTD method discretizes the above governing equation with respect to space and time. do. Specifically, when the spatial discretization widths in the x, y, and z directions are Δx, Δy, and Δz for sound pressure and the time discretization width is Δt, respectively, the particle velocity is Δx / 2, for space. Discretize Δy / 2, Δz / 2 and time by shifting from the sound pressure reference point by Δt / 2.

When the above governing equations are differentially approximated with such discretization, the governing equations of equations (1) to (4) are transformed as shown in the following equations.

i, j, and k are spatial steps indicating the number of sound pressure reference points in the x, y, and z directions in the analysis space, respectively. n is a time step indicating the number of reference points in terms of time. For example, un ^{+ 1} (i + 1/2, j, k) indicates the value of the particle velocity in the x direction when the spatial step is i + 1/2, j, k and the time step is n + 1, and p ^{n + 1. / 2} (i, j, k) indicates the value of sound pressure when the spatial step is i, j, k and the time step is n + 1/2.

By alternately calculating the particle velocity represented by the equations (5) to (7) and the sound pressure represented by the equation (8) for the analysis space (entire space step), the state in the new time step is calculated one after another. Can be (Leap Frog algorithm).

With reference to FIG. 8, the condition setting unit 330 sets the calculation conditions of the computer simulation using the acoustic FDTD method in the acoustic FDTD calculation unit 320. Specifically, the condition setting unit 330 includes the sound pressure p and bulk modulus κ used in the governing equation, the spatial discretization width Δx, Δy, Δz, the time discretization width Δt, and the diaphragm 220 which is a sound source (FIG. 2). ), The number of time steps, and other constants such as the speed of sound are set in the acoustic FDTD calculation unit 320. The housing 210 (FIG. 2) is formed of a rigid body, and the sound wave is set to be totally reflected.

Regarding the input to the diaphragm 220, in this embodiment, the Gaussian pulse represented by the following equation is given to the diaphragm 220, and the acoustic characteristic data when the diaphragm 220 is excited by the Gaussian pulse is obtained. It is calculated by the acoustic FDTD calculation unit 320.

g (t) = A · exp [− {(nΔt−T) /0.29T} ² ]… (9)
T = 0.646 / fc ... (10)
In the above, A is the maximum amplitude value, and fc is the frequency at which the frequency spectrum of the amplitude of the Gaussian pulse is reduced by 3 dB.

The acoustic FDTD calculation unit 320 uses a computer simulation using the acoustic FDTD method for each of a large number of randomly generated design parameters received from the design parameter setting unit 310 according to the calculation conditions set by the condition setting unit 330. The sound pressure and particle velocity in the analysis space (the space within the boundary 260 of the small speaker 200 shown in FIGS. 2 and 3) are sequentially calculated. Then, the acoustic FDTD calculation unit 320 acquires the time response data of the sound pressure at the analysis point P0 shown in FIG. 2 for each design parameter, and performs a fast Fourier transform (FFT) process or the like on the time response data. The acoustic characteristic data (data of the sound pressure frequency characteristic and the phase frequency characteristic at the analysis point P0) is calculated by performing the above.

The acoustic characteristic data calculated in this way constitutes learning data by pairing with the design parameters corresponding to the acoustic characteristic data, and corresponds to a large number of design parameters generated by the design parameter setting unit 310. The training data 110-1 to 110-N of the above is stored in the database 120 (FIG. 7).

The spatial discretization widths Δx, Δy, Δz (hereinafter, the grid having this spatial discretization width is referred to as a “grid”) set by the condition setting unit 330 is described in the analysis space (FIGS. 2 and 3). In the space (space within the boundary 260 of the small speaker 200 shown), it is preferable to appropriately change the size (roughness) of the grid according to the region.

FIG. 11 is a diagram illustrating an example of the concept of grid setting by the condition setting unit 330. FIG. 11 is a cross-sectional view of the small speaker 200, in which a schematic grid is superimposed on the cross-sectional view of FIG.

With reference to FIG. 11, the condition setting unit 330 includes a region 410 in the vicinity of the diaphragm 220 which is a sound source, a region 420 in the housing 210 excluding the region 410, and a region 430 outside the housing 210 (acoustic propagation space 250). Then, change the size (roughness) of the grid and set it. Specifically, the size of the grid in the region 410 including the diaphragm 220 is relatively small, and the size of the grid in the region 430 (acoustic propagation space 250) outside the housing 210 is relatively large. The grid is set.

In this way, the accuracy of computer simulation by the acoustic FDTD method can be improved and the calculation time can be improved by appropriately changing the size of the grid according to the region, instead of making the size of the grid uniform in the analysis space. Can be shortened.

The three-stage grid division as shown in FIG. 11 is an example, and the grid division method is not limited to that shown in FIG. Further, the example shown in FIG. 11 is schematically shown for explaining the difference in relative size between the three stages of grids, and does not show the absolute size of each grid. do not have.

With reference to FIG. 7 again, the DNN learning unit 130 includes a CPU, a ROM for storing a processing program, a RAM for temporarily storing data, and the like. Then, the DNN learning unit 130 learns the DNN using a large number of learning data 110-1 to 110-N stored in the database 120, and calculates the design parameters of the small speaker 200 from the acoustic characteristic data. A trained model for operating a computer (computing device 20 in FIG. 1) is generated. That is, this trained model is actually the DNN itself in which the causal relationship between the acoustic characteristic data and the design parameters is learned using the training data 110-1 to 110-N stored in the database 120.

FIG. 12 is a configuration diagram of a DNN used in the trained model generated by the DNN learning unit 130. With reference to FIG. 12, the DNN includes an input layer, an output layer, and a plurality of intermediate layers. By providing a plurality of intermediate layers, the DNN has higher generalization performance than the conventional neural network (NN), and is particularly useful for a model having a complicated shape (design parameter) such as a small speaker 200. Is.

The data input to the input layer is weighted at the joint portion with the first intermediate layer and input to the first intermediate layer. Each unit (neuron) in the first intermediate layer is provided with an activation function, and the output of the activation function of each unit is in the connection portion with the next second intermediate layer (not shown). It is weighted and input to the second intermediate layer. Such data propagation is sequentially performed in the plurality of intermediate layers, and finally the data is output to the output layer.

The output data of the output layer is compared with the teacher data. Then, the gradient to the input layer is calculated based on the error between the output data of the output layer and the teacher data, and the weighting (parameter) between each unit is updated using the calculated gradient (error backpropagation method). .. By performing such processing with a large amount of training data, the causal relationship between input and output is modeled (generation of trained model).

FIG. 13 is a diagram showing input / output of DNN in the present embodiment. With reference to FIG. 13, the acoustic characteristic data of the learning data received from the database 120 (FIG. 7) is input to the input layer of the DNN. In this embodiment, the profile of the frequency characteristic of the absolute value (Pa) of the sound pressure and the profile of the frequency characteristic of the phase are used as the acoustic characteristic data. As an example, each profile of the frequency characteristics of sound pressure and phase is discretized in a predetermined frequency width Δf and given to the input layer of the DNN.

Data corresponding to the design parameters is output to the output layer. That is, the number of units in the output layer is the same as the number of variables of the design variable vector X = [LCx, LHx, LHy, LHz, Hymin, Hzmin] (6 in this embodiment). Then, the output data of the output layer is compared with the design parameter (teacher data) paired with the acoustic characteristic data of the learning data input to the input layer, and the weighting (parameter) between each unit is based on the comparison error. Is updated. The DNN learning unit 130 generates a trained model by performing the above processing on a large number of learning data stored in the database 120 (FIG. 7).

With reference to FIG. 7 again, the learning condition setting unit 140 sets the learning conditions for generating the trained model in the DNN learning unit 130. For example, the learning condition setting unit 140 sets the number of intermediate layers of the DNN (composed of four layers in this embodiment) and the number of units (number of neurons) of each of the input layer, each intermediate layer, and the output layer. .. Further, for example, the learning condition setting unit 140 sets a condition for ending learning by each learning data. Specifically, in order to prevent overfitting of DNN by the training data, test data (design parameters and acoustic characteristic data) different from the training data are prepared. Then, in learning with each learning data, test data (acoustic characteristic data) is given to the input layer of the DNN each time learning is performed, and an error (test error) between the output data of the output layer and the test data (design parameter). The above-mentioned predetermined number of times is set as a condition for ending the learning by each learning data so that the learning of the learning data is completed if the minimum value of is not updated continuously a predetermined number of times (early end).

FIG. 14 is a flowchart illustrating a procedure for generating a trained model used in the arithmetic unit 20 shown in FIG. With reference to FIGS. 7 and 8 together with FIG. 14, first, a large number of learning data 110-1 to 110-N are generated in the processes of steps S10 to S40. That is, in the generation system 300, the design parameter setting unit 310 sets the design parameters for generating the learning data (step S10). The design parameter setting unit 310 considers the symmetry of the design parameters as described in FIG. 9, and randomly sets the design parameters as described in FIG.

Next, the acoustic FDTD calculation unit 320 calculates acoustic characteristic data corresponding to the design parameters set in step S10 by computer simulation using the above-mentioned acoustic FDTD method (step S20). Then, the learning data for pairing the design parameter set in step S10 and the acoustic characteristic data calculated in step S20 based on the design parameter is stored in the database 120 of the learning system 100 (step S30). ..

Next, the acoustic FDTD calculation unit 320 determines whether or not the number of learning data stored in the database 120 is a predetermined number N or more (step S40). The predetermined number N is, for example, a value of several thousand levels. When it is determined in step S40 that the number of training data is less than the predetermined number N (NO in step S40), the process is returned to step S10, considering the symmetry of the design parameters, and randomly, the next Design parameters are set.

When it is determined in step S40 that the number of training data is a predetermined number N or more (YES in step S40), a trained model is subsequently generated in the processes from steps S50 to S70. That is, in the learning system 100, the learning condition setting unit 140 sets the DNN learning conditions in the DNN learning unit 130 (step S50). As described above, as the learning conditions, for example, the configuration of the DNN (the number of layers, the number of units (the number of neurons), etc.), the conditions for terminating the learning with each learning data, and the like are set.

Next, in the DNN learning unit 130, the causal relationship between the acoustic characteristic data and the design parameters is learned by the DNN using a large number of learning data 110-1 to 110-N stored in the database 120 (step). S60). Specifically, the acoustic characteristic data of the learning data is given to the input layer of the DNN, and a large number of learning data 110-1 to 110 are used as the teacher data using the design parameters of the learning data paired with the acoustic characteristic data. DNN learning is performed using −N.

Then, when the learning of the DNN by the learning data 110-1 to 110-N is completed, the learned model (actually, the structure of the DNN and the parameter indicating the weighting between the units) is output from the DNN learning unit 130. (Step S70).

FIG. 15 is a flowchart illustrating a procedure of processing executed by the arithmetic unit 20 shown in FIG. With reference to FIG. 15, the arithmetic apparatus 20 obtains acoustic characteristic data (profiles of sound pressure (Pa) and phase frequency characteristics as shown in FIGS. 5 and 6) input from the input device 10 by the user. Obtained from the input device 10 (step S110). The acoustic characteristic data is not the acoustic characteristic data of the learning data stored in the database 120 of the learning system 100, but the desired acoustic characteristic data input from the input device 10 by the user who is about to design the small speaker. ..

Next, the arithmetic unit 20 inputs the acoustic characteristic data acquired in step S10 into the trained model (post-learning DNN) generated in advance according to the procedure shown in FIG. 14, and uses the trained model to design parameters ( LCx, LHx, LHy, LHz, Hymin, Hzmin) shown in FIGS. 2 and 3) are calculated (step S120).

Then, the arithmetic unit 20 outputs the calculated design parameters LCx, LHx, LHy, LHz, Hymin, and Hzmin to the display device 30 (step S130). In the display device 30, the numerical value of the design parameter received from the arithmetic unit 20 may be displayed as it is, or the structure of the small speaker 200 having such a design parameter may be illustrated.

As described above, in this embodiment, input is performed using a trained model for operating the computer (computing device 20) so as to output the design parameters of the small speaker based on the acoustic characteristic data of the small speaker. Design parameters are calculated from the desired acoustic characteristic data input by the device 10. The trained model is composed of DNNs trained to calculate design parameters from acoustic characteristic data. This makes it possible to design a small speaker without relying on a cut-and-try design by an engineer.

A plurality of learning data 110-1 to 110-N used for DNN learning, each including a design parameter and acoustic characteristic data of a small speaker having the design parameter, are computer simulations using the acoustic FDTD method. Generated by. As a result, a large number of learning data 110-1 to 110-N used for learning DNN are prepared without actually creating various acoustic structures and actually measuring them in order to prepare a plurality of learning data. be able to.

As described above, according to the design support device 1 for the small speaker of the present disclosure, various acoustic structures are actually created in order to prepare the learning data of DNN without relying on the cut and try design by the engineer. Since it is not necessary to carry out actual measurement, the time and cost required for designing a small speaker can be significantly reduced.

Further, in this embodiment, the size (roughness) of the grid in the computer simulation using the acoustic FDTD method is appropriately changed depending on the region. Specifically, the grid in the region near the diaphragm 220, which is a sound source, is made small, and the grid in the acoustic propagation space 250 outside the housing is made large (FIG. 11). As a result, the accuracy of the acoustic FDTD method can be improved and the calculation time can be shortened as compared with the case where the grid size is uniform.

Further, according to this embodiment, the acoustic FDTD calculation unit 320 calculates acoustic characteristic data for each design parameter by limiting the analysis space in consideration of the symmetry of the shape of the design target (small speaker 200). The calculation load and calculation time can be reduced.

Further, according to this embodiment, since the DNN is trained using a large number of training data based on the randomly changed design parameters, the training becomes a local solution when the trained model is generated by the DNN. It is possible to avoid falling and improve the accuracy of the trained model.

Regarding the training data 110-1 to 110-N used for generating the trained model, in the above embodiment, of the acoustic characteristic data given to the input layer of the DNN at the time of training by the DNN learning unit 130. The sound pressure frequency characteristic is a profile of the frequency characteristic of the absolute value (Pa) of the sound pressure as shown in FIG. 16, but instead of the profile of the frequency characteristic of the absolute value (Pa) of the sound pressure, FIG. 17 shows. A profile of the frequency characteristics of the sound pressure level (dB) as shown may be given to the input layer of the DNN. The profile of the frequency characteristic of the sound pressure level (dB) shows the characteristics at frequencies other than the resonance frequency fr better than the profile of the frequency characteristic of the absolute value (Pa) of the sound pressure, and emphasizes the profile of the frequency characteristic. When designing, the accuracy of the trained model can be improved.

Further, when the trained model is generated, the profile of the frequency characteristic of the absolute value (Pa) of the sound pressure is given to the input layer of the DNN for learning, or the profile of the frequency characteristic of the sound pressure level (dB) is given to the DNN. The user may be able to select whether to give the sound to the input layer for learning. For example, referring to FIG. 7 again, information indicating the user's selection result is set in the learning condition setting unit 140 to the DNN learning unit 130, and learning is performed using the profile of the frequency characteristic of the sound pressure level (dB). When this is selected by the user, the frequency characteristic data of the absolute value (Pa) of the sound pressure received from the database 120 may be converted into decibels (dB) and given to the input layer of the DNN.

When a trained model is generated by giving a profile of the frequency characteristic of the sound pressure level (dB) to the input layer of the DNN, the desired acoustic characteristic given to the trained model by the arithmetic unit 20 of the design support device 1. The sound pressure frequency characteristic in the data also needs to be a profile of the frequency characteristic of the sound pressure level (dB).

Further, in the above embodiment, the acoustic characteristic data is composed of data indicating the frequency characteristic of the sound pressure and data indicating the frequency characteristic of the phase, but data indicating the frequency characteristic of the sound pressure. May be only. By using not only the frequency characteristics of sound pressure but also the frequency characteristics of phase as in the above-described embodiment, it is expected that the learning accuracy of DNN and the design accuracy of a small speaker using the trained model will be improved. can.

The embodiments disclosed this time should be considered to be exemplary in all respects and not restrictive. The scope of the present disclosure is shown by the claims rather than the description of the embodiments described above, and is intended to include all modifications within the meaning and scope equivalent to the claims.

1 Design support device, 10 input device, 20 arithmetic device, 30 display device, 100 learning system, 110-1, 110-2, ..., 110-N learning data, 120 database, 130 DNN learning unit, 140 learning Condition setting unit, 200 small speaker, 210 housing, 220 diaphragm, 230 sound hole, 240 air chamber, 250 acoustic propagation space, 260 boundary, 270, 410-430 area, 300 generation system, 310 design parameter setting unit, 320 Acoustic FDTD calculation unit, 330 condition setting unit.

Claims (8)

It is a design support device for small speakers.
An input device for the user to input acoustic characteristic data indicating the acoustic characteristics of the small speaker, and
The design parameters are calculated from the acoustic characteristic data input by the input device using a trained model for operating the computer to output the design parameters of the small speaker based on the acoustic characteristic data. Equipped with a configured arithmetic unit
The trained model uses a plurality of training data, each including the design parameter and the acoustic characteristic data of a small speaker having the design parameter, generated by computer simulation using the acoustic FDTD method. A design support device for a small speaker configured by a deep neural network learned to calculate the design parameters from acoustic characteristic data. The small speaker is configured to include a diaphragm serving as a sound source and a space inside the housing formed inside the housing of the small speaker.
The design support for a small speaker according to claim 1, wherein the size of the divided grid in the computer simulation using the acoustic FDTD method is smaller in the order of the diaphragm, the space inside the housing, and the acoustic propagation space outside the housing. Device. The design support device for a small speaker according to claim 1 or 2, wherein the design parameters included in each of the plurality of learning data are randomly generated within a predetermined range. The small speaker design support device according to any one of claims 1 to 3, wherein the acoustic characteristic data includes data indicating a frequency characteristic of sound pressure at a predetermined analysis point. For the acoustic characteristic data used as the plurality of learning data, either data indicating the frequency characteristic of the absolute value of the sound pressure or data indicating the frequency characteristic of the sound pressure level is selected according to the request of the user. , The design support device for a small speaker according to claim 4. The design support device for a small speaker according to claim 4 or 5, wherein the acoustic characteristic data further includes data indicating a frequency characteristic of the phase at the analysis point. The design support device for a small speaker according to any one of claims 1 to 6, wherein the design parameter includes the size and position of a sound hole formed in the small speaker and the size of an air chamber. It is a speaker design support method.
A step of generating a trained model for the computer to function to output the design parameters of the speaker based on the acoustic characteristic data indicating the acoustic characteristics of the speaker.
The step of inputting the acoustic characteristic data by the user and
Including the step of calculating the design parameter from the input acoustic characteristic data using the trained model.
The step of generating the trained model is
A step of generating a plurality of learning data including the design parameter and the acoustic characteristic data of the speaker having the design parameter, respectively, by computer simulation using the acoustic FDTD method.
A speaker design support method including a step of generating the trained model by training a deep neural network so as to calculate the design parameter from the acoustic characteristic data using the plurality of learning data.

Images