JP6621706B2 - Sound source detection device - Google Patents

Description

  The present invention relates to a sound source detection apparatus that detects a sound source that is a cause of sound generated from a machine.

  In a machine having a large number of drive units such as a general machine and a transport machine, it is not easy to specify a sound source (sound source) even if an abnormal sound or the like is generated therefrom. Therefore, conventionally, as a sound source detection device, a device to which a technique for specifying a sound source of a sound such as an abnormal sound is applied is known. As a calculation algorithm of such a technique, a beam forming method and an acoustic intensity method are well known. In the beam forming method, a spatial sound pressure distribution is calculated, and in the sound intensity method, sound intensity (sound intensity and direction) is calculated. The beam forming method and the sound intensity method are different from each other in calculation content, and both are long and short, so both are often used together. For example, Patent Document 1 describes an example of a sound source detection device using such a technique.

  The sound source detection device described in Patent Document 1 is a device that obtains a high-resolution sound pressure distribution using a small number of microphones such as four microphones employed in the sound intensity method. Specifically, sound signals obtained by receiving sound by three or four sound sensors, for example, four sound sensors arranged one by one at four vertices of a regular tetrahedron, are input, and the input sound Based on the signal, the sound pressure distribution in the sound source plane including the sound source is obtained by beam forming calculation using the minimum dispersion method (Capon method), the direction of sound arrival is searched, and the search result is presented.

JP2015-219138A

  According to the sound source detection device described in Patent Document 1 described above, the portion where sound is generated is identified from the surface of the machine by visualizing the portion where sound is generated according to the intensity of the sound. be able to.

  However, even if the part where the sound is generated is found from the surface of the machine, the part is not necessarily the sound source that is the factor that generates the sound. For example, a machine has many vibration parts that are likely to be sound sources and has high rigidity and easy transmission of vibrations. Therefore, it is not uncommon for a part where sound is generated to be away from the sound source. For this reason, even if the part where the sound is generated is specified, it is often necessary to search for the sound source of the sound from that part.

  The present invention has been made in view of such circumstances, and an object thereof is to provide a sound source detection apparatus capable of detecting a sound source of a sound generated from a machine with high accuracy.

  A sound source detection apparatus that solves the above problem is a sound source detection apparatus that detects a sound source of a sound generated from a detection target having a plurality of vibration sources, and is a time series of target sound that is a sound generated from the detection target. A storage unit that stores data and sound or vibration time-series data output from each of the plurality of vibration sources; and a predetermined period length of the time-series data of the target sound and the time-series data of the sound or vibration A coherence calculation unit that calculates coherence for each period shifted so that a part of a predetermined period length overlaps and obtains time-series coherence, and the coherence calculation unit includes the time-series coherence for each vibration source. A first calculation unit that calculates the vibration source related to the target sound, and each sound corresponding to one or a plurality of vibration sources selected by the first calculation unit or The time series coherence between one time series data obtained by superimposing dynamic time series data and the time series data of the target sound is calculated, and the selected one or more vibration sources constitute the target sound. A time series coherence between the second calculation unit for determining the time and the time series data of the target sound for each sound or vibration time series data corresponding to the one or more vibration sources determined by the second calculation unit. And a third calculation unit that calculates the degree of contribution to each vibration source with respect to the target sound.

  Since the target sound obtained from the machine as the detection target is generated based on multiple factors, which of the machine's vibration sources contributes to the generation of the target sound to a high degree. Not easy to understand. In this respect, according to such a configuration, since the degree of contribution of each vibration source to the target sound is calculated, it can be said that the vibration source has a high degree of contribution to the target sound generated from the machine, that is, the sound source of the target sound. The vibration source can be detected with high accuracy. For example, it is possible to specify a vibration source located in a part away from the target sound generation part as a sound source. Then, necessary countermeasures such as calming can be performed on the vibration source contributing to the identified target sound.

  Moreover, since the sound source of the target sound is a vibration source that is a vibrating part of the machine, the target sound can be compared with the generated vibration instead of the sound generated by the vibration source. That is, the coherence between the target sound and the vibration source can be calculated based on any relationship between “sound” and “sound” and “sound” and “vibration”. Furthermore, since the time series coherence of sound or vibration is calculated for a predetermined period length, if the sound required for identification or the length of vibration is about the predetermined period length, the contribution also to relatively short sounds, etc. The degree of can be calculated. This also makes it possible to detect with high accuracy a vibration source (sound source) that contributes greatly to the target sound generated from the machine.

As a preferred configuration, the coherence calculation unit sets the predetermined period length to a length that allows the vibration source to be identified from time series data of sound or vibration of the vibration source.
The length of sound or vibration time-series data required for specifying the vibration source differs depending on the characteristics of the sound and vibration. Thus, as in this configuration, by setting the predetermined period length to a predetermined period in which the vibration source can be specified, the possibility that the vibration source is specified as the sound source is increased.

As a preferred configuration, the coherence calculation unit sets the predetermined period length to the length of the shortest period among the lengths that can identify each of the one or more vibration sources.
For each vibration source, the length of time series data of sound or vibration necessary for specifying the vibration source is different. Therefore, as in this configuration, the number of vibration sources to be detected can be increased by setting the predetermined period length to the vibration source with the shortest time-series data length necessary for identification.

As a preferred configuration, the sound and vibration frequencies include an audible frequency and an inaudible frequency.
According to such a configuration, it becomes possible to specify the sound source for any sound or vibration of a frequency or sound that can be heard and for a sound or vibration of a frequency that cannot be heard.

As a preferred configuration, the sound source detection device holds time series data of the vibration source in the storage unit prior to measurement of the target sound.
According to such a configuration, it is possible to acquire the time series data of the vibration source under more appropriate conditions by holding the time series data of the vibration source in advance. In addition, an increase in processing load at the same time as the target sound detection process is suppressed.

As a preferred configuration, the sound source detection device collects time series data of the vibration source and time series data of the target sound at the same timing.
According to such a configuration, the time required for measurement can be shortened by acquiring the time series data of the target sound and the vibration source at the same timing. In addition, it is expected that the relationship between the acquired time series data of the target sound and the time series data of the vibration source is suitably maintained.

  According to the sound source detection device, the sound source of the sound generated from the machine can be detected with high accuracy.

The block diagram which shows the schematic structure about one Embodiment which actualized the sound source detection apparatus. The block diagram which shows schematic structure which collects the time series data of the sound or vibration corresponding to a sound source in the embodiment. The block diagram which shows schematic structure of the sound source detection apparatus in the embodiment. In the embodiment, the schematic diagram which shows typically the relationship between the sound (target sound) which generate | occur | produced from the detection target object, and the sound of each sound source contained in the sound. The figure which shows the example of the coherence calculated conventionally. The figure for demonstrating time-sequential coherence in the said one Embodiment. 5 is a flowchart showing a processing procedure of sound source detection processing of the sound source detection device in the embodiment. In the embodiment, it is a figure which shows the example of the sound for demonstrating coherence, Comprising: (a)-(d) is a figure which each shows the time change of four sounds from which an amplitude differs each typically. The figure which shows typically the example of the time change of the target sound which the four sounds shown in FIG. 8 overlap and generate | occur | produce in the same embodiment. In the embodiment, (a)-(d) is a figure which shows each frequency response of the four sounds shown in FIG. The figure which shows the multi-coherence of the sound which piled up the four sounds shown in FIG. 8, and the object sound in the same embodiment. In the embodiment, (a)-(d) is a figure which shows the partial coherence of each of four sounds shown in FIG. 8, and an object sound. In the same embodiment, it is a figure which shows the time change waveform and frequency characteristic of a sound, Comprising: (a) is a figure corresponding to a 1st sound source, (b) is a figure corresponding to a 2nd sound source, (c). The figure corresponding to a 3rd sound source, (d) is a figure corresponding to one object sound. In the same embodiment, it is a figure which each shows the coherence of an object sound and the sound from each sound source, (a) is a figure which shows the coherence with the sound from a 1st sound source, (b) is a figure from a 2nd sound source. The figure which shows the coherence with a sound, (c) is a figure which shows the coherence with the sound from a 3rd sound source. The figure which shows multi-coherence with the sound in which the target sound and two selected sounds overlap in the embodiment. In the same embodiment, it is a figure which shows the partial coherence of an object sound and each selected sound, Comprising: (a) is a figure which shows the contribution degree of the sound of the 1st sound source in an object sound, (b) is in an object sound. The figure which shows the contribution degree of the sound of a 2nd sound source.

With reference to FIGS. 1-16, one Embodiment of a sound source detection apparatus is described.
As shown in FIG. 1, the sound source detection device 10 of the present embodiment measures the sound generated from the automobile 1 as the target sound N when the door 1A of the automobile 1 as a detection target is closed, and the measurement. A sound source that is a generation source of the target sound N is detected.

First, it will be described how the time series data collection device 3 shown in FIG. 2 collects the time series data 5 (see FIG. 3) of the sound of the portion used for sound source detection by the sound source detection device 10.
The time-series data collection device 3 shown in FIG. 2 collects time-series data 5A of the door latch sound N1, time-series data 5B of the door panel radiated sound N3, and time-series data 5C of the door-to-frame hit sound N2. . Then, the time series data collected by the time series data collection device 3 is held in the sound source detection device 10 prior to the sound source detection process.

  More specifically, the time series data collection device 3 is connected to sensor bodies 4A, 4B, and 4C that detect sound and vibration, and acquires sound and vibration signals from the sensor bodies 4A, 4B, and 4C as time series data. To do. That is, each sensor body 4A, 4B, 4C includes a sound sensor and a vibration sensor. The sensor body 4A is disposed in the vicinity of the “door latch”, the sensor body 4C is disposed in the vicinity of the “door panel”, and the sensor body 4B is disposed in the vicinity of “the door and the frame hit”. Therefore, the time-series data collection device 3 acquires the time-series data 5A including the door latch sound N1 and the vibration related to the door latch via the sensor body 4A. Similarly, the time-series data collection device 3 acquires time-series data 5B including the radiated sound N3 of the door panel and vibration related to the radiation of the door panel via the sensor body 4C, and the door and frame via the sensor body 4B. The time series data 5C including the hit sound N2 and the vibration related to the door and frame hit are acquired. Each time-series data is acquired by closing a door or generating a door closing sound for each part alone. Further, the time length of the time series data 5A of the door latching sound N1 is at least a time length capable of identifying the door latching sound N1 from the time series data 5A. Similarly, the length of the time series data 5B of the door panel radiated sound N3 is at least a time length that allows the door panel radiated sound N3 to be identified from the time series data 5B. The time length of the series data 5C is at least a time length that can be identified from the time series data 5C as the door-to-frame hit sound N2. Each time series data 5A, 5B, 5C may include only one time series data used in sound source detection processing among sound and vibration.

  The sound source detection device 10 shown in FIGS. 1 and 3 measures the target sound N generated from the automobile 1 via the sound pressure sensor 2 and stores time-series data 12 for the measured target sound N. Further, the sound source detection device 10 holds time-series data 5 of the sound of the portion that is a candidate for the sound source of the target sound N. The time series data 5 of the sound of the part includes the time series data 5A of the door latch sound N1 collected by the time series data collection device 3, the time series data 5B of the radiated sound N3 of the door panel, and the door and frame. The time series data 5C of the hit sound N2 is included. The sound source detection device 10 calculates a coherence function between the measured target sound N, the door latch sound N1 which is the sound of three parts, the door panel radiated sound N3, and the door and frame hit sound N2. Based on this, the correlation height (contribution) of the sound of each sound source to the target sound N is calculated.

  As shown in FIG. 4, in the present embodiment, the target sound N changes in time series as shown in the graph L10. The target sound N includes a door latch sound N1 whose amplitude changes in time series shown in the graph L11, a door panel radiated sound N3 whose amplitude changes in time series shown in the graph L12, and a time series change in which the amplitude shows in the graph L13. This is a sound in which the hitting sound N2 of the door and the frame is superimposed. If the sound source of the target sound has a different factor from the above or is estimated to be a different factor, the sensor body is disposed at a position corresponding to a factor different from the above. Further, if there are more than three sound sources or candidates, more than three sensor bodies may be provided.

  In FIG. 4, the target sound N is highly correlated with the door latching sound N1, the door-to-frame hitting sound N2, and the door panel radiating sound N3 between times t0 and t4. On the other hand, when the period is divided finely, it is also shown that the level of correlation (degree of contribution) of the door latch sound N1, the door-to-frame hit sound N2, and the door panel radiated sound N3 with respect to the target sound N is different. Specifically, during the period from time t0 to time t1, the correlation of the door latching sound N1 is particularly high, and during the period from time t1 to time t2, the correlation of the frame hitting sound N2 is particularly high. Further, the correlation between the door and frame hit sound N2 and the door panel radiated sound N3 is particularly high during the period from time t2 to time t3, and the correlation between the door panel radiated sound N3 is particularly high during the period from time t3 to time t4. . Therefore, the correlation height with respect to the sound source varies depending on the setting of the period for calculating the correlation height for the target sound N and the length of data used for the calculation of the coherence function. In other words, the detection accuracy of a sound source having a high correlation can be changed by setting a period for calculating the level of correlation in the target sound N and setting the length of data used for calculating the coherence function.

  A sound source detection apparatus 10 shown in FIG. 3 is configured by a microcomputer including a CPU, a ROM, a RAM, and the like. The sound source detection device 10 executes various processes in the sound source detection device 10 by executing various programs held in, for example, a ROM or RAM by the CPU. In the present embodiment, the sound source detection device 10 performs a sound source detection process on the time series data 12 of the target sound obtained by measuring the sound generated from the automobile 1 as the sound source detection process.

  The sound source detection device 10 includes the time series data 5A of the door latching sound N1, the time series data 5B of the radiated sound N3 of the door panel, and the time series data 5C of the door-to-frame hitting sound N2 as described above. A storage unit 11 that holds data 5 is provided. Further, the storage unit 11 holds time-series data 12 of the target sound N acquired through the sound pressure sensor 2.

The sound source detection device 10 includes a coherence calculation unit 15 that calculates a coherence function that is an index indicating the level of correlation of the sound source with the target sound N.
The coherence calculation unit 15 performs a multi-point input multi-point output (MIMO: Multiple Input and Multiple Output) analysis. MIMO analysis is an analysis method for obtaining a frequency response function and coherence function for simultaneous inputs by providing multiple excitation points when obtaining vibration characteristics such as the natural vibration frequency and vibration shape of the structure in the experimental mode analysis of the structure. It is. In the present embodiment, the coherence calculator 15 can perform MIMO analysis even when at least one of the input and the output is one point. In the present embodiment, the coherence calculation unit 15 performs a coherence function calculation, particularly a time-series coherence calculation. In addition, the coherence calculator 15 calculates three types of coherence for the time-series coherence. Therefore, in the following, the calculation of the coherence function, the calculation of the time series coherence, and the calculation of the three types of coherence will be described.

First, the outline of the calculation of the coherence function will be described.
The coherence function γ ² indicates the degree of correlation between the input and output of the system, and a value between “0” and “1” is obtained as a calculation result. Here, the calculation result of the coherence function γ ² (f) is obtained as the ratio of the height at which the system output correlates with the system input with respect to the frequency f. For example, when the coherence function γ ² (f) is “1”, it is indicated that all the outputs of the system are caused by the inputs of the system at the frequency f (high correlation). Further, when the coherence function γ ² (f) is “0”, for the frequency f, it is indicated that the output of the system is completely unrelated to the input of the system (no correlation). When “0 <γ ² (f) <1”, it is indicated that the system output includes a signal unrelated to the system input. Signals irrelevant to system inputs include unknown inputs, noise generated within the system, system nonlinearity, or system time delay.

The coherence function γ ² is expressed by the following formula (1). However, the input complex spectrum obtained by Fourier transform is “X”, and the obtained output complex spectrum is “Y”. The power spectrum on the input side obtained by multiplying “X” by the conjugate complex number is “Wxx”, the power spectrum on the output side obtained by multiplying “Y” by the conjugate complex number is “Wyy”, and one signal (“X”). The cross spectrum obtained by multiplying the complex conjugate of the complex spectrum of the second signal by the complex spectrum of another signal (“Y”) is defined as “Wxy”.

That is, the coherence function γ ² is obtained by dividing the square of the absolute value of the cross spectrum Wxy by the power spectra Wxx and Wyy of the input and output of the system, and is determined by two of the input and the output. Indicates the frequency component of power. Then, by averaging the values of the plurality of coherence functions γ ² calculated for different frames by the number existing in the period during which the coherence is calculated, a coherence value indicating the degree of correlation is obtained. For example, in the present embodiment, the power spectrum Wyy is set based on the target sound N, and the power spectrum Wxx is set based on one or more sounds selected as inputs from the sounds N1 to N3 of the sound source. The number of frames to be averaged is a predetermined number constituting one data of time series coherence, for example, “5”.

  For example, FIG. 5 shows an example of coherence when sound is emitted from one sound source. When the vibration of the sound source is indicated by a graph L50 and the intensity of the measured sound is indicated by a graph L51, the calculated coherence is indicated by a graph L52. The graph L52 indicates that the darker the color, the higher the coherence value, that is, the higher the correlation. That is, it is shown that the closer the time ta0 is, the higher the correlation of the vibration of the sound source with respect to each frequency of the sound. Then, the correlation decreases as the vibration of the sound source decreases, but again, the sound increases as the sound source vibrates at time ta1, and the correlation of the vibration of the sound source increases with respect to each frequency of the sound. It has been shown. Normally, the coherence function is averaged even over time. For example, when the calculation is started from time ta0, the correlation at time ta0 becomes relatively high because the correlation is based on the time average in the period close to time ta0, which is the period in which the sound source is vibrating. On the other hand, the correlation at time ta1 is kept low because it is based on the time average from time ta0 to time ta1 including the period in which the sound source is not vibrating, and the correlation at time ta2 is the period in which the sound source is not vibrating. Since it is based on the time average from time ta0 to time ta2 that is included for a long time, it is kept lower.

Next, an outline of time series coherence calculation will be described.
An overview of time-series coherence will be described with reference to FIG. In time-series coherence, the time length of data used for the calculation of the coherence function is determined as a frame length LF as a predetermined period length, and a plurality of calculation timings are provided at intervals shorter than the frame length LF. At each calculation timing, the calculation of the coherence function during the frame length LF is averaged to obtain time-series coherence. That is, the time-series coherence is calculated at intervals shorter than the frame length LF, for example, in FIG. 6, five calculation timings are provided in the frame length LF. And the calculation result of coherence is shown as the graph L57. The graph L57 indicates that the correlation is high when the vibration portion of the graph L55 is included in the frame length LF, and indicates that the correlation is low when the vibration portion of the graph L55 is not included.

The coherence calculator 15 of this embodiment will be described in detail.
The coherence calculation unit 15 illustrated in FIG. 3 calculates the coherence at the frame length LF for each calculation timing for the time series data of the target sound N and the time series data of the sounds N1 to N3 of each sound source. At this time, the coherence calculation unit 15 calculates coherence for each of the overlapping frame lengths, and sets these as time-series coherence. The interval between the two calculation timings is a predetermined shift time. For example, if the frame length is 100 [ms] and the shift time is 20 [ms], the two coherences are periods in which 80 [ms] overlap each other while 20 [ms] do not overlap each other. Note that the frame length may be selected from preset values, or a more stable value may be found and set in the process of calculating coherence.

Hereinafter, the coherence in the present embodiment is assumed to be time-series coherence.
As shown in FIGS. 3 and 7, the coherence calculator 15 includes a first calculator 20 that calculates normal coherence, a second calculator 30 that calculates multi-coherence, and a third calculator that calculates partial coherence. 40. In the present embodiment, when “normal coherence”, “multi-coherence”, and “partial coherence” are calculated by the coherence calculation unit 15, the combinations of the sounds N1 to N3 of the sound source as input are different, or the coherence is The frequency range to be calculated is different.

  First, an overview of the three types of coherence described above will be described with reference to FIGS. 8A to 8D show temporal changes of four uncorrelated input signals L20 to L23 as inputs, and FIG. 9 shows an output that is an output based on the four input signals L20 to L23. The time change of the signal L30 is shown. The four input signals L20 to L23 have an amplitude of "1.4 times" when the amplitude of the input signal L20 of FIG. 8A is "1". The amplitude of the input signal L22 in FIG. 8C is “2 times”, and the amplitude of the input signal L23 in FIG. 8D is “0.5 times”. The output signal L30 is about “4 times” where the amplitude is large. Further, FIG. 8 and FIG. 9 show frequency signals with substantially constant amplitude, but for convenience of illustration, they are shown as signals with a constant width.

  Of the three types of coherence, “normal coherence” is calculated as a high correlation between a certain input and output in a plurality of uncorrelated input sources. If the correlation between one input and one output is high, the value is close to “1”, and the value is smaller than “1” when influenced by other than the input.

  As shown in FIGS. 10A to 10D, in “normal coherence”, the level of correlation between the input signal L20 and the output signal L30 is calculated (graph L40), and the input signal L21 is output to the output signal L30. Is calculated (graph L41). Further, the height of the correlation between the input signal L22 and the output signal L30 is calculated (graph L42), and the height of the correlation between the input signal L23 and the output signal L30 is calculated (graph L43). According to this calculation process, an input signal having a high correlation with the output signal L30 can be selected. In other words, an input signal that is not correlated with the output signal L30 can be excluded. That is, even when a large number of sound sources whose correlation with the output is unknown are input by this calculation process, sound sources having high correlation with the output from those inputs are appropriately selected.

  In this embodiment, the first calculation unit 20 of the coherence calculation unit 15 calculates “normal coherence” to determine whether or not each input of a plurality of inputs has a correlation with the output. Initially, the input (sound source) that has a correlation with the output is estimated, but its authenticity is unknown, but this calculation process identifies the input unrelated to the output, and the identified unrelated input is detected as the sound source. Can be excluded from processing.

  “Multi-coherence” is a so-called multiple relevance function, and the degree of correlation (degree of contribution) between all inputs and one output is calculated. If the correlation between all the inputs and one output is high, the value is close to “1”, and if affected by other than the input, the value is smaller than “1”.

  As shown in FIGS. 8, 9, and 11, the height of the correlation with the output signal L30, which is a signal obtained by superimposing four uncorrelated input signals L20 to L23 at the same time, is calculated (graph L44). ). By this calculation process, it is determined whether or not each of the four uncorrelated input signals L20 to L23 has a high correlation with the output signal L30. For example, if the sound source for the output is unknown, the input may be insufficient. Therefore, in this calculation process, if “1”, it indicates that all inputs correlated with the output are selected, and if it is smaller than “1”, there is a possibility that some sound sources are not input. Is shown.

  In the present embodiment, it is determined whether or not the output is satisfied with all the selected inputs. If it is determined in this calculation process that the output is satisfied by all the inputs, it is indicated that the inputs are appropriately selected without being insufficient. On the other hand, if it is determined that all the inputs do not satisfy the output, it indicates that the selected input is insufficient. Then, by combining the two “normal coherence” and the “multi-coherence”, the input irrelevant to the output is excluded, and it is possible to determine whether there is an insufficient input. .

  Further, “partial coherence” is a so-called partial relevance function, and a high correlation between only one input and one output among multi-point inputs is calculated. Again, if the correlation between one input and one output is high, the value is close to “1”, and if affected by other than the input, the value is smaller than “1”.

  As shown in FIGS. 12A to 12D, the correlation height to the output signal L30 is calculated independently for each of the four uncorrelated input signals L20 to L23. For example, the correlation level of the input signal L20 is about “0.15” (graph L45), the correlation level of the input signal L21 is about “0.3” (graph L46), and the correlation level of the input signal L22 is high. Is about “0.5” (graph L47), and the correlation level of the input signal L23 is less than “0.05” (graph L48). That is, the sum of the correlation heights calculated for the uncorrelated input signals L20 to L23 is approximately “1” (= 0.15 + 0.3 + 0.5 + 0.05). In other words, the sum of the correlation heights of the uncorrelated inputs is calculated as the correlation height for the outputs of all the inputs. Therefore, the correlation height of each uncorrelated sound source is obtained as a ratio to the whole.

  As described above, in the present embodiment, since the sound source selected without excess or deficiency in the two calculation processes of “normal coherence” and “multi-coherence” is input, the correlation of these inputs to the output is high. 1 ”. Therefore, the degree of correlation of each input to the output can also be obtained as the degree of contribution to the output.

  Furthermore, by narrowing down the frequency range for calculating “partial coherence” for the output of one input, or by narrowing down the time range, it is possible to narrow down the frequency range and time range with high correlation among the outputs. Become.

  Next, an example of the operation of the sound source detection process performed by the sound source detection device 10 will be described with reference to FIG. Prior to the sound source detection process being performed, the sound source detection device 10 holds data to be input. Further, at this time, there is a feature that the target sound that is a sound that needs to detect the sound source in a prior investigation or the like exists in the target range KN including the time T and the frequency range F shown in FIG. It is assumed that it is known. The target sound is, for example, fluctuating sound, noise, or anxious sound.

  As shown in FIGS. 13A to 13C, the sound source detection apparatus 10 holds data to be input before the sound source detection process is performed. At this time, it is assumed that time series data of the sound of the first portion is obtained as a time change graph L60 shown in FIG. This graph L60 has a characteristic in the frequency range F corresponding to the same range M60 as the target range KN, and shows that this characteristic lasts for the time T. That is, the sound of the first portion is considered to have high coherence in the range M60 with respect to the characteristic of the sound to be worried about. At this time, the time range and frequency range deviating from the target range KN are not calculated. Also, it is assumed that the time-series data of the sound of the second part is obtained as a time change graph L61 shown in FIG. This graph L61 has a characteristic in the frequency range F corresponding to the same range M61 as the target range KN, and shows that this characteristic lasts for the time T. That is, the sound of the second portion is also considered to have high coherence in the range M61 with respect to the characteristic of the sound that is of interest. Furthermore, it is assumed that the time-series data of the sound of the third portion is obtained as a time change graph L62 shown in FIG. This graph L62 is characterized in that the sound of the third portion is characterized by a time range and a frequency range that are outside the target range KN.

  As shown in FIG. 7, the sound source detection device 10 acquires time-series data of the target sound (step S10 in FIG. 7). In step S10, it is assumed that the time-series data of the input target sound is obtained as a time change graph L63 shown in FIG. This graph L63 has a characteristic in the frequency range F corresponding to the target range KN, and this characteristic lasts for the time T. At this time, the target sound graph L63 includes a section “A” corresponding to the graphs L61 and L62, and a section “B” corresponding to the graph L63. However, the section “B” is not shown because the time range and frequency range of the frequency characteristics are out of the target range KN.

  Subsequently, as illustrated in FIG. 7, in the sound source detection device 10, the first calculation unit 20 of the coherence calculation unit 15 performs normal coherence calculation (step S <b> 11 in FIG. 7), and the second calculation unit 30 performs multicoherence. The calculation (step S12 in FIG. 7) is performed, and the third calculation unit 40 calculates partial coherence (step S13 in FIG. 7).

Among these, as shown in FIGS. 7 and 14, “normal coherence” is calculated in step S11.
When the input time-series data is represented by a graph L60, the first calculator 20 calculates the level of correlation of the target sound that is the output with the time-series data (graph L63). Here, the characteristic time and frequency characteristic range range M60 of the graph L60 overlaps with the characteristic time and frequency characteristic range target range KN of the graph L63, and it is calculated that the correlation is high (FIG. 14A). reference). Similarly, when the input time-series data is a graph L61, the first calculator 20 calculates the level of correlation of the target sound that is the output to the time-series data (graph L63). Here, the characteristic time and frequency characteristic range range M61 of the graph L61 overlaps the characteristic time and frequency characteristic range target range KN of the graph L63, and it is calculated that the correlation is high (FIG. 14B). reference). Further, the first calculation unit 20 calculates the level of correlation of the target sound, which is an output when the input time-series data is the graph L62, with the time-series data (graph L63). Here, it is calculated that the characteristic time and frequency characteristic range range of the graph L62 and the characteristic time and frequency characteristic range target range KN of the graph L63 do not overlap and the correlation is low (FIG. 14 (c). )reference).

  Based on the calculation of “normal coherence” in step S11 shown in FIG. 7, it is determined that the correlation of the target sound with the time series data (graph L63) is high in graph L60 and graph L61, and low in graph L62. Is done. Therefore, in the subsequent processing, the sound source detection processing is performed on the graphs L60 and L61 with high correlation, and the low-correlation graph L62 such as irrelevant can be omitted from the processing. Whether the correlation is high or low is determined by comparing the calculated “normal coherence” value with the first threshold value. The first threshold value is a threshold value used for determining whether or not there is a correlation, and a value suitable for the determination is set from “0” to “1”.

  Then, as shown in FIGS. 7 and 15, in step S <b> 12, “multi-coherence” is calculated based on the input determined to have a high correlation in the calculation of “normal coherence” in step S <b> 11.

  That is, the second calculation unit 30 correlates the time-series data of the sound in which the graph L60 and the graph L61, which are determined to be highly correlated by the first calculation unit 20, with the time-series data of the target sound (graph L63). Calculate the height of. Here, it is calculated that the characteristic time and frequency characteristic range range M65 of the time series data of the overlapping sound overlaps the characteristic time and frequency characteristic range target range KN of the graph L63, and the correlation is high ( FIG. 15). In other words, the level of correlation of all input time-series data to the time-series data of the target sound (graph L63), excluding time-series data irrelevant to the target sound, is calculated from the input. If the value of the correlation height is close to “1” as a result of such calculation, it is confirmed that the input composed of the graph L60 and the graph L61 is the sound source of the graph L63. On the other hand, if the value of the correlation height is “0” as a calculation result, it is confirmed that the input composed of the graph L60 and the graph L61 is not the sound source of the graph L63. Further, if the value of the correlation height is less than “1” and greater than “0” as a calculation result, it is indicated that the sound source of the graph L63 may include sounds other than the graph L60 and the graph L61. Whether the correlation is high or low is determined by comparing the calculated “multicoherence” value with a second threshold value. The second threshold value is a threshold value used for determining whether the graph L63 is composed of the selected sound source, and a value suitable for the determination is set from “0” to “1”.

  That is, in the sound source detection process shown in FIG. 7, whether or not the selected input is appropriate as the sound source of the target sound by excluding unnecessary inputs in step S11 and confirming that there is no shortage of inputs in step S12. Determine whether or not. If the input is appropriate as a sound source, the process of the next step S13 is performed. Conversely, if the input is inappropriate as a sound source, the sound source detection process is stopped. In this case, the portion estimated as the sound source of the target sound may be reselected.

  Finally, as shown in FIGS. 7 and 16, in step S13, the time series data (graph L63) of the target sound for each of the graph L60 and the graph L61 determined as the sound source of the target sound N in step S12. Calculate “partial coherence”. At this time, for example, the correlation height is calculated for each of the first frequency range f1 and the second frequency range f2 obtained by dividing the characteristic frequency range F into two.

  When it is determined that the input selected by the second calculation unit 30 is appropriate as an output sound source, the third calculation unit 40 first sets the first frequency range f1 and the second frequency when the input is a graph L60. The height of the correlation to the time series data (graph L63) of the target sound in the frequency range f2 is calculated. At this time, for example, when the graph L60 is in the first frequency range f1, it is determined that the correlation is high in the time and frequency range M66, and conversely, in the second frequency range f2, it is determined that the correlation is not high. (See FIG. 16A). Further, the third calculating unit 40 determines that the correlation is high in the time and frequency range M67 when the graph L61 is in the second frequency range f2, and conversely, when the graph L61 is in the first frequency range f1, the correlation is high. It is determined that there is not (see FIG. 16B). Whether the correlation is high or low is determined by comparing the calculated “partial coherence” value with the third threshold value. The third threshold value is a threshold value used for determining whether the correlation of the sound source is high or low for all or part of the target sound, and a value suitable for the determination is set from “0” to “1”. .

  Accordingly, the sound source detection device detects that the feature of the graph L60 is highly correlated with the time T and the frequency range M66 in the time-series data (graph L63) of the target sound, and the feature of the graph L61 is the time T and the frequency. It is detected that the correlation is high in the range M67.

  In this way, for the input that is appropriately narrowed down through the calculation of “normal coherence” and “multi-coherence”, “partial coherence” is calculated for the output, so that the calculated coherence has a correlation with the target sound. Is reflected with high accuracy. If the accuracy of the high correlation with the target sound is high, it is easy to take appropriate measures against the sound source, and the utility value of the sound source detection processing by the sound source detection device is increased.

As described above, according to the sound source detection device according to the present embodiment, the following effects can be obtained.
(1) Since the target sound N obtained from the automobile as the detection target is generated based on multiple factors, which vibration source among the automobile vibration sources contributes to the generation of the target sound to a high degree. It is not easy to see if it is. In this respect, in this embodiment, since the level of correlation (degree of contribution) of each sound source with respect to the target sound N is calculated, a sound source having a high degree of contribution to the target sound N generated from the automobile 1 can be obtained with high accuracy. Can be detected. For example, even if the sound source is in a portion away from the portion where the target sound N is generated, it can be specified as the sound source. Then, necessary countermeasures such as calming can be performed on the sound source contributing to the identified target sound N.

  In addition, since time-series coherence is calculated for a predetermined frame length (period length), if the sound required for identification or the length of vibration is about the predetermined frame length, the contribution to relatively short sounds, etc. The degree of can be calculated. This also makes it possible to detect a sound source that has a high contribution to the target sound N generated from the automobile 1 with high accuracy.

  (2) The length of sound time-series data necessary for specifying a sound source differs depending on the characteristics of sound and vibration. Therefore, by setting the frame length to a predetermined period in which the vibration source can be specified, the possibility that the vibration source is specified as the sound source is increased.

  (3) By holding the time series data 5A to 5C of the sound source in advance, it becomes possible to acquire the time series data of the vibration source under more appropriate conditions. In addition, an increase in processing load at the same time as the detection process of the target sound N is suppressed.

(Other embodiments)
In addition, the said embodiment can also be implemented with the following aspects.
In the above embodiment, the case where the time series data of the sound of the sound source is collected in advance is illustrated, but the present invention is not limited to this, and the time series data of the sound of the sound source may be performed simultaneously with the collection of the target sound. By acquiring the time series data of the target sound and the sound source at the same timing, the time required for measurement can be shortened. In addition, it is expected that the relationship between the acquired time series data of the target sound and the time series data of the sound source is suitably maintained.

  -In the said embodiment, although the frequency of a sound is not prescribed | regulated in particular, the frequency of a sound may be an audible frequency, a non-audible frequency, and those may be mixed. As a result, not only the sound that can be heard but also the sound that cannot be heard, the sound source can be specified for those sounds.

  In the above embodiment, the frame length may be selected from preset values, or a more stable value may be found and set in the process of calculating coherence. In the above embodiment, the latter case is illustrated. Further, the length of sound or time-series data required for specifying the vibration source differs for each vibration source. Therefore, the frame length may be set to the vibration source having the shortest time-series data length necessary for identification. As a result, the number of vibration sources to be detected as sound sources can be increased.

  In the above embodiment, the case where the time series data of the sound and vibration of the sound source is collected by the time series data collection device 3 is illustrated, but the present invention is not limited to this, and the time series data of the sound and vibration of the sound source is collected by the sound source detection device. May be collected. For example, a sensor body may be connected to the sound source detection device, and the sound source detection device may collect sound and vibration time-series data of the vibration source.

  In the above-described embodiment, the case of collecting time-series data of sound and vibration of a sound source has been illustrated. However, the present invention is not limited thereto, and only one of sound or vibration time-series data of a vibration source serving as a sound source is collected. May be. Correlation can be obtained even if the target sound is changed to the sound emitted from the vibration source and compared with the vibration generated simultaneously. That is, the coherence between the target sound and the vibration source can be calculated based on any relationship between “sound” and “sound” and “sound” and “vibration”.

  -In above-mentioned embodiment, each sensor body 4A-4C illustrated about the case provided with the sound sensor and the vibration sensor. However, the present invention is not limited to this, and the sensor body may include only a sound sensor or a vibration sensor. In addition, as a plurality of sensor bodies, those having only a sound sensor, those having a vibration sensor, those having a sound sensor and those having a vibration sensor are mixed. It may be. Even in this case, as described above, the correlation can be obtained by comparing the target sound with the time-series data of either the sound or vibration of the sound source acquired by the sensor body.

  In the above-described embodiment, the case where one part of sound is acquired as one time-series data has been illustrated. However, the present invention is not limited to this, and two or more portions of sound may be acquired as one time-series data. In this case, when it is determined that there is a high correlation in one time-series data consisting of two or more parts of the target sound, the target part of the time-series data is reduced by reducing the target part of the time-series data. You may make it narrow down.

  -In the above-mentioned embodiment, although the case where there was one object sound was illustrated, it is not restricted to this but a plurality of object sounds may be sufficient. It is only necessary to appropriately calculate various coherences between the respective target sounds and component sounds.

  In the above embodiment, the case where the sound source is detected with respect to the sound when the door of the automobile is closed is illustrated. However, not limited to this, the sound that is detected by the sound source is generated in the car, such as the sound around the engine of the car, the sound when idling the car, the sound when opening and closing the hood and trunk, and the sound when opening the door. Any sound may be used.

  In the above embodiment, the case where the sound source is detected for the sound generated from the automobile is illustrated. However, the present invention is not limited to this, and the object that generates sound may be a general machine or a transport machine having a plurality of sound sources.

  DESCRIPTION OF SYMBOLS 1 ... Automobile, 1A ... Door, 2 ... Sound pressure sensor, 4A-4C ... Sensor body, 10 ... Sound source detection apparatus, 11 ... Memory | storage part, 15 ... Coherence calculation part, 20 ... 1st calculation part, 30 ... 2nd calculation Part, 40 ... 3rd calculation part.

Claims (6)

A sound source detection device for detecting a sound source of a sound generated from a detection target having a plurality of vibration sources,
A storage unit that stores time-series data of a target sound that is a sound generated from the detection target object, and time-series data of a sound or vibration output from each of the plurality of vibration sources;
As the time series coherence, the coherence in the predetermined period length of the time series data of the target sound and the time series data of the sound or vibration is calculated for each period shifted so that a part of the predetermined period length overlaps. A coherence calculator to obtain,
The coherence calculation unit calculates the time-series coherence for each vibration source and selects the vibration source related to the target sound, and the one or more selected by the first calculation unit The time series coherence between one time series data obtained by superimposing the time series data of each sound or vibration corresponding to the vibration source and the time series data of the target sound is calculated, and the selected one or more vibration sources are A second calculation unit that determines that the target sound is configured; and a sound corresponding to one or a plurality of vibration sources determined by the second calculation unit or the time of the target sound for each time-series data of vibration. A sound source detection apparatus comprising: a third calculation unit that calculates time series coherence with the sequence data and calculates the degree of contribution to each vibration source with respect to the target sound. The sound source detection device according to claim 1, wherein the coherence calculation unit sets the predetermined period length to a length capable of specifying the vibration source from sound or vibration time-series data of the vibration source. The sound source detection device according to claim 2, wherein the coherence calculation unit sets the predetermined period length to a length of a shortest period among lengths that can identify each of the one or the plurality of vibration sources. The sound source detection device according to claim 1, wherein the sound and vibration frequencies include an audible frequency and a non-audible frequency. The sound source detection device according to any one of claims 1 to 4, wherein the sound source detection device holds time-series data of the vibration source in the storage unit prior to measurement of the target sound. The sound source detection device according to any one of claims 1 to 4, wherein the sound source detection device collects time series data of the vibration source and time series data of the target sound at the same timing.