JP2015175772A - Sound identification condition setting support device and sound identification condition setting support method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To alleviate a processing load and accelerate the process when performing an allophone feature extraction by computer in a spectrogram of a picked-up sound.SOLUTION: When a sound identification condition setting support device divides a predetermined frequency range of spectrogram data for an allophone sample sound with allophone and no allophone sample sound without allophone into a plurality of blocks, and generates an identification function by use of the histogram data for each block, a sound strength of which average order of degree among the histogram data of the allophone sample sounds in the same frequency band is high is selected as target strength, and an identification function is generated by using only the target strength. A plurality of identification functions are generated while changing number of strength selected as the target strength so as to select the identification function having the highest identification accuracy rate among them.

Description

  The present invention relates to a sound identification condition setting support device and a sound identification condition setting support method that support the setting of an identification condition for identifying the presence or absence of an abnormal sound in a collected sound.

  For example, when considering measures to remove abnormal noise emitted from power transmission lines, or when diagnosing equipment malfunctions based on abnormal noise emitted from equipment, of the sounds collected near the power transmission line or equipment. It is determined by using a pattern recognition method whether or not there is an abnormal sound. Specifically, in many cases, the collected sound is subjected to spectrum analysis, and the presence or absence of the abnormal sound in the collected sound is identified based on the characteristics of the collected sound or the frequency of the abnormal sound (see Patent Document 1).

  However, if the frequency or intensity of the abnormal sound changes over time, it may be difficult to accurately identify the presence or absence of the abnormal noise based solely on the static characteristics of the sampled sound or the abnormal frequency. is there. In such a case, it is desirable to analyze the spectrogram of the sampled sound and identify the presence or absence of the abnormal sound in the sampled sound in consideration of the temporal change in the frequency or intensity of the sampled sound or abnormal sound. Thereby, even when the frequency of abnormal noise or the intensity of sound changes with time, the presence or absence of abnormal noise can be identified with high accuracy.

JP-A-11-241945

  In order to identify the presence or absence of an abnormal sound in the collected sound by the pattern recognition method using the spectrogram of the collected sound, it is necessary to extract the characteristic of the abnormal sound expressed as the spectrogram. The superiority or inferiority of the feature extraction greatly affects the identification accuracy.

  There is no known general or unified method for extracting abnormal sound features. 2. Description of the Related Art Conventionally, an operator performs feature extraction of an abnormal sound by a method of displaying a spectrogram of a collected sound on a computer monitor and finding the characteristic of the spectrogram of the abnormal sound by visually observing the spectrogram.

  However, it takes a long time to find the characteristics of the abnormal sound that enables high-accuracy identification of the abnormal sound by looking at the spectrogram displayed on the computer monitor, which imposes a heavy burden on the operator. In addition, since it is often dependent on the experience of the operator whether or not the feature of the abnormal sound that enables high-precision identification of abnormal noise can be found, high-precision identification is performed when there is no skilled worker. It is difficult to realize.

  Therefore, development of a method for automatically extracting features of abnormal sounds by a computer is now underway. If abnormal noise feature extraction can be performed automatically by a computer, abnormal noise feature extraction can be performed easily and reliably, and the burden on the operator regarding the processing can be reduced.

  However, in order to extract abnormal noise features that enable highly accurate identification of abnormal noise, a large number of samples of collected sound must be processed by a computer, and the amount of data to be processed is enormous. May be. For this reason, it is difficult to perform the feature extraction process for abnormal sounds quickly with a general personal computer. To perform the feature extraction process for abnormal sounds quickly, an expensive and large specialized computer must be prepared. There must be.

  The present invention has been made in view of, for example, the above-described problems. The object of the present invention is to reduce the burden of processing on a computer when performing feature extraction of abnormal sounds or processing corresponding thereto in a spectrogram of collected sound. An object of the present invention is to provide a sound identification condition setting support device and a sound identification condition setting support method that can be reduced and speed up the processing.

  In order to solve the above-described problem, a first sound identification condition setting support device according to the present invention is a sound identification condition setting support device that supports setting of an identification condition for identifying the presence or absence of an abnormal sound. Receiving the setting noise spectrogram data indicating the spectrograms of the plurality of setting noise sample sounds including each, dividing the predetermined frequency range into a plurality of blocks in each setting noise spectrogram data, Anomalous noise histogram generator for setting for generating anomalous noise histogram for setting indicating a histogram of sound intensity distribution in each block, and anomalous noise histogram data for setting is generated by the anomalous noise histogram generator for setting In the histogram data of the blocks in the same frequency band in the plurality of setting different sound spectrogram data set , Calculate the average or cumulative value of the ranks of the frequency of each sound intensity, specify the intensity of a predetermined number of sounds in descending order of the average value or cumulative value, and determine the intensity of these specified sounds. A spectrogram for generation showing a spectrogram of a plurality of sample sounds for generation including an intensity sampler for generation including abnormal sound and a non-noise sample sound for generation including no abnormal sound A generation histogram generation unit that receives data, divides a predetermined frequency range into a plurality of blocks in each of the generation spectrogram data, and generates generation histogram data indicating a histogram of a sound intensity distribution in each block; For identifying the presence or absence of abnormal noise using the frequency of the target intensity in each generation histogram data A discriminant function generator for generating another function, and a verification sample showing spectrograms of a plurality of verification sample sounds including a verification abnormal sound sample sound including abnormal sound and a verification non-sound sample sound including no abnormal sound A verification histogram generator that receives spectrogram data, divides a predetermined frequency range into a plurality of blocks in each verification spectrogram data, and generates verification histogram data indicating a histogram of sound intensity distribution in each block; , The frequency of the target intensity in each verification histogram data is input to the discrimination function to identify the presence / absence of abnormal noise, and the discrimination accuracy rate, which is the probability that the discrimination function correctly identifies the presence / absence of abnormal noise, And a correct rate calculation unit for calculating.

  Further, the second sound identification condition setting support device of the present invention is the above-described first sound identification condition setting support device of the present invention described above, while changing the setting of the selection number, the intensity selection unit, the generation A process by the histogram generation unit, the discrimination function generation unit, the verification histogram generation unit, and the accuracy rate calculation unit is executed a plurality of times, and the process is executed a plurality of times. An identification condition selecting unit that selects the identification function having the highest classification accuracy rate among the identification functions and the target strength selected when generating the identification function as the identification conditions is provided.

  Further, the third sound identification condition setting support device of the present invention changes the setting of the number of blocks and the selection number when dividing a predetermined frequency range in each spectrogram data into a plurality of blocks, The processing by the noise histogram generation unit, the intensity selection unit, the generation histogram generation unit, the discrimination function generation unit, the verification histogram generation unit, and the accuracy rate calculation unit is executed a plurality of times, and the processing is performed a plurality of times. As a result of execution, the discrimination function having the highest discrimination accuracy rate among the plurality of discrimination functions generated by the discrimination function generation unit, the number of blocks set when generating the discrimination function, and the discrimination function are generated. An identification condition selection unit is provided that selects the target intensity selected when performing the identification as the identification condition.

  The first sound identification condition setting support method of the present invention is a sound identification condition setting support method that supports setting of an identification condition for identifying the presence or absence of an abnormal sound, and is used for a plurality of settings including abnormal sound. The setting specific sound spectrogram data indicating the spectrogram of the abnormal sound sample is received, the predetermined frequency range is divided into a plurality of blocks in each setting specific sound spectrogram data, and the sound intensity in each block A setting noise histogram generation step for generating setting noise histogram data indicating a distribution histogram, and the plurality of settings in which the setting noise histogram data is generated in the setting noise histogram generation step In the histogram data of the block of the same frequency band in the abnormal sound spectrogram data for each, Calculate the average or cumulative value of the rank of the frequency of the intensity, specify the intensity of the selected number of sounds in descending order of the average value or cumulative value, and select the specified sound intensity as the target intensity Receiving a spectrogram data for generation indicating a spectrogram of each of a plurality of generation sample sounds including a generation noise sample sound including generation noise and a generation non-sound sample sound including no generation noise, In each of the generation spectrogram data, a generation histogram generation step of dividing a predetermined frequency range into a plurality of blocks and generating generation histogram data indicating a histogram of a sound intensity distribution in each block; Using the frequency of the target intensity in the histogram data, an identification function for identifying the presence or absence of abnormal noise And a spectrogram data for verification showing a spectrogram of a plurality of sample sounds for verification including a sample sound for verification including abnormal sounds and a non-noise sample sound for verification that does not include abnormal sound. Receiving each verification spectrogram data, dividing a predetermined frequency range into a plurality of blocks, and generating a verification histogram data indicating a histogram of a sound intensity distribution in each block; The correct answer for inputting the frequency of the target intensity in the verification histogram data to the discrimination function to identify the presence / absence of abnormal noise, and calculating the discrimination accuracy rate that is the probability that the discrimination function correctly identifies the presence / absence of abnormal noise A rate calculation step, changing the setting of the number of selections, the strength selection step, and the generating hiss A plurality of processes generated in the discrimination function generation process are performed as a result of executing the process by the program generation process, the discrimination function generation process, the verification histogram generation process, and the accuracy rate calculation process a plurality of times. The discriminant function having the highest discrimination accuracy rate among discriminant functions and the target strength selected when generating the discriminant function are selected as the discriminating conditions.

  Further, the second sound identification condition setting support method of the present invention is the above-described first sound identification condition setting support method of the present invention described above, wherein the predetermined frequency range in each spectrogram data is divided into a plurality of blocks. While changing the setting of the number of blocks and the number of selections, the setting noise histogram generation step, the intensity selection step, the generation histogram generation step, the discrimination function generation step, the verification histogram generation step, and the correct answer As a result of executing the process by the rate calculation step a plurality of times and executing the processing a plurality of times, the discrimination function having the highest discrimination accuracy rate among the plurality of discrimination functions generated in the discrimination function generation step, and the discrimination function are generated As the identification condition, the number of blocks set when performing and the target intensity selected when generating the identification function Characterized in that it constant.

  According to the present invention, when a computer performs a feature extraction of an abnormal sound or a process corresponding thereto in a spectrogram of a collected sound, the processing burden on the computer can be reduced, and the process can be speeded up.

It is a block diagram which shows the sound identification condition setting assistance apparatus by embodiment of this invention. It is a flowchart which shows the sound identification condition setting assistance process by the sound identification condition setting assistance apparatus by embodiment of this invention. It is a flowchart which shows the sound identification condition setting assistance process following FIG. It is a flowchart which shows the sound identification condition setting assistance process following FIG. It is a wave form diagram of a strange sound sample sound. It is a graph which shows the spectrogram of an unusual sound sample sound. It is a graph which shows the histogram of the intensity distribution of the sound in one block of the spectrogram of the unusual sound sample. It is a graph which shows the histogram of the intensity distribution of the sound in other one block of the spectrogram of an unusual sound sample sound. It is a graph which shows the histogram of the intensity distribution of the sound in another one block of the spectrogram of a strange sound sample sound. It is explanatory drawing which shows the state which divided | segmented the spectrogram of the sample sound into the block of 20 blocks. It is explanatory drawing which shows the state which divided | segmented the spectrogram of the sample sound into the block of the number of 30 blocks. It is explanatory drawing which shows the maximum frequency intensity | strength etc. of each frequency band in the spectrogram data of one unusual sound sample. It is explanatory drawing which shows the count value etc. of the candidate block of the same frequency band in the spectrogram data of a noise sample sound. It is explanatory drawing which shows the order | rank of the intensity | strength of the sound in one histogram data of a strange sound sample sound. It is explanatory drawing which shows the average value of the intensity | strength order | rank of the sound intensity in the several histogram data of a noise sample sound. It is explanatory drawing which shows the calculation method of the identification correct rate of an identification function. It is explanatory drawing which shows the discrimination | determination accuracy rate of the discrimination function corresponding to each setting of the number of blocks, the selection number c of a frequency band, and the selection number d of a sound intensity. It is a block diagram which shows the abnormal sound determination apparatus by embodiment of this invention. It is a flowchart which shows the noise determination process by the noise determination apparatus by embodiment of this invention.

  Embodiments of the present invention will be described below with reference to the drawings.

(Outline of processing by the sound identification condition setting support device)
FIG. 1 shows a sound identification condition setting support apparatus according to an embodiment of the present invention. In FIG. 1, a sound identification condition setting support apparatus 1 according to an embodiment of the present invention is an apparatus that supports the setting of an identification condition for identifying the presence or absence of abnormal sounds in a collected sound. The collected sound is, for example, a sound collected at a place near a power transmission line or equipment. The abnormal noise is, for example, a sound in which people living around the power transmission line complain of discomfort or a sound that does not occur when the device is normal.

  The process for supporting the setting of the identification condition performed by the sound identification condition setting support device 1 (sound identification condition setting support process) roughly generates a discrimination function for identifying the presence or absence of abnormal sounds in the collected sound. Or a process to select. This process is substantially equivalent to a process for extracting features of abnormal sounds.

  The sound identification condition setting support device 1 generally performs the following process as the sound identification condition setting support process. First, as preparation for processing, an operator prepares audio data of abnormal sound samples and audio data of non-natural sound samples. The abnormal sound sample sound is a sample of the collected sound including the abnormal sound. An allophone sample sound is a sample of a sampled sound that does not contain an allophone. The audio data is digital data having an audio file format such as WAV, MP3 (MPEG Audio Layer-3), AAC (Advanced Audio Coding), and the like.

  For example, the operator prepares a plurality of voice data of abnormal sample sounds and a plurality of voice data of non-sound sample sounds by the following method. In other words, the operator collects a plurality of sounds by executing recording at the same place, reproduces the collected sounds sequentially, listens to them, and determines the presence or absence of abnormal sounds in each sound. Then, the operator classifies the audio data corresponding to the sound including the abnormal sound into the sound data of the abnormal sound sample sound, and the sound data corresponding to the sound not including the abnormal sound is sampled without the abnormal sound. Classify into sound data.

  Next, the worker inputs sound data of a plurality of abnormal sound samples and sound data of a plurality of sound samples without sound to the sound identification condition setting support device 1. The sound identification condition setting support device 1 generates spectrogram data indicating a spectrogram of speech for each of the input speech data.

  Subsequently, the sound identification condition setting support device 1 (1) selection of the target frequency band, (2) selection of the target intensity, (3) generation of the discrimination function, (4) calculation of the discrimination accuracy rate of the discrimination function, ( 5) Select a deterministic identification function. The outline of the processes (1) to (5) is as follows.

(1) Selection of target frequency band First, the process of selecting a target frequency band is as follows. That is, the sound identification condition setting support device 1 generates a discrimination function using the abnormal sound sample and the non-noise sample sound, and the discrimination function is generated using the abnormal sound sample and the non-noise sound. This is not performed using the entire frequency band of the sample sound, but only using a limited frequency band that is expected to include the frequency of the abnormal sound. In order to realize this, the sound identification condition setting support device 1 specifies a frequency band that is expected to include an abnormal noise frequency, and selects the frequency band as a “target frequency band”.

  Specifically, the sound identification condition setting support device 1 first sets the “number of blocks” according to a predetermined rule. Subsequently, the sound identification condition setting support device 1 divides the predetermined frequency range into the number of blocks indicated by the number of blocks for each spectrogram data of the abnormal sound sample, and the sound intensity distribution for each of the blocks Histogram data indicating the histogram of the current is generated. The predetermined frequency range is, for example, a range from about 20 Hz to about 20000 Hz, which is a general audible range.

  Subsequently, the sound identification condition setting support device 1 selects, as the “candidate block”, a block whose sound intensity has increased despite the frequency being increased for each spectrogram data of the abnormal sound sample. That is, in the collected sound that does not include the abnormal sound, it is generally considered that the intensity of the sound decreases as the frequency increases. On the other hand, in the sampled sound including the abnormal sound, the intensity of the frequency of the abnormal sound is larger than the intensity of the frequencies before and after the abnormal sound. Therefore, if there is a frequency band in which the intensity of the sound increases in the sampled sound even though the frequency has increased, the sampled sound contains the abnormal sound, and the noise is included in the frequency band. It is highly possible that the frequency is included. Therefore, the sound identification condition setting support device 1 selects, as a candidate block, a block in a frequency band in which the sound intensity increases even though the frequency is increased.

  Here, the selection of candidate blocks will be described more specifically. In the histogram data of the sound intensity distribution, if the intensity of the sound having the maximum frequency is referred to as “maximum frequency intensity”, the sound identification condition setting support apparatus 1 has the highest frequency in each spectrogram data of the sampled sound with abnormal sound. For all the blocks except the low band block, compare the maximum frequency intensity of the histogram data of the two adjacent blocks, and the maximum frequency intensity of the histogram data of the block with the higher frequency band is the lower frequency band When the maximum frequency intensity of the histogram data of the block is larger, the block with the higher frequency band is selected as a candidate block.

  Subsequently, the sound identification condition setting support device 1 sets the “number of selections c” according to a predetermined rule. Subsequently, the sound identification condition setting support device 1 counts the number of candidate blocks in the same frequency band for each of the frequency bands of the blocks in all the spectrogram data of the generated abnormal sound samples. Subsequently, the sound identification condition setting support device 1 specifies the number of frequency bands indicated by the selection number c in descending order of the count value, and sets the specified frequency band and the adjacent frequency band lower than the frequency band as the target frequency. Select as a belt.

  In other words, each of the abnormal sound samples includes an abnormal sound having the same frequency, and it is considered that the abnormal sound is a target to be identified. Among the spectrogram data of abnormal sample sounds generated by the sound identification condition setting support device 1, when many spectrogram data include candidate blocks of the same frequency band, the frequency bands of the candidate blocks are included in the frequency bands of the candidate blocks. There is a high possibility that an abnormal noise frequency is included. Further, the intensity distribution of the abnormal noise may be characterized by the difference between the intensity distribution of the frequency band including the frequency of the abnormal noise and the intensity distribution of the adjacent frequency band lower than the frequency band. Based on such a concept, the sound identification condition setting support device 1 includes a frequency band in which a relatively large number of candidate blocks exist in the predetermined frequency range in the generated spectrogram data of a plurality of abnormal sound samples. The adjacent frequency band lower than the frequency band is selected as the target frequency band.

(2) Selection of target strength Next, processing for selecting target strength is generally as follows. That is, the sound identification condition setting support device 1 divides a predetermined frequency range into the number of blocks indicated by the set number of blocks for each of spectrogram data of abnormal sound samples and non-noise sample sounds. Histogram data showing a histogram of the sound intensity distribution is generated for each of the above, and the discrimination function is generated using these histogram data, but the discrimination function is generated using the entire intensity range of the histogram data. Instead, the discriminant function is generated using only a limited intensity range in which the frequency is expected to increase due to the presence of abnormal noise. In order to realize this, the sound identification condition setting support device 1 specifies an intensity range in which the frequency is expected to increase due to the presence of an abnormal sound, and selects the intensity range as the “target intensity”.

  In other words, in the sampled abnormal sound, if a sound in a limited intensity range is generated with a high frequency in a limited frequency band, the sound is likely to be abnormal. Therefore, in the histogram data of the target frequency band in the spectrogram data of the abnormal sound sample sound, if there is a part where the intensity frequency is significantly larger than the surrounding area, that part is the part where the frequency is increased due to the presence of the abnormal sound Is likely to hit. Based on such an idea, the sound identification condition setting support device 1 has a relatively high intensity (for example, the rank of the frequency in the histogram data of the same target frequency band in the spectrogram data of the abnormal sound sample sound). The first, second, and third strengths) are selected as target strengths.

  Specifically, the sound identification condition setting support device 1 sets the “number of selections d” according to a predetermined rule. Subsequently, the sound identification condition setting support device 1 calculates the average rank of the frequency of each sound in the histogram data of the blocks of the same target frequency band in all the spectrogram data of the generated abnormal sound samples. The intensity of the number of sounds indicated by the selection number d is specified in descending order of the average rank, and the intensity of the specified sounds is selected as the target intensity.

(3) Generation of Discriminant Function Next, discriminant function generation processing is generally as follows. That is, the sound identification condition setting support device 1 includes a part of a plurality of input abnormal sound samples, for example, half of the abnormal sound samples and one of the input non-natural sound samples. A discriminant function is generated using a part, for example, half of all sound samples. Specifically, the sound identification condition setting support device 1 includes spectrogram data of half of the different sampled sounds out of the plurality of input sampled sounds, and a plurality of input non-sampled sounds. For each of the spectrogram data of half of the non-sound sample sound, the predetermined frequency range is divided into the number of blocks indicated by the set number of blocks, and only the blocks in the target frequency band among the divided blocks Histogram data indicating a histogram of sound intensity distribution is generated. If the histogram data has already been generated for each block of the spectrogram data of the input abnormal sound sample in order to perform the above-described target frequency band selection processing, the discrimination function generation processing is performed. Therefore, it is not necessary to newly generate the histogram data, and the histogram data of the target frequency band may be selected and used from the histogram data generated to perform the process of selecting the target frequency band.

  Subsequently, the sound identification condition setting support device 1 generates an identification function using only the frequency of the target intensity in the generated histogram data of the target frequency band.

(4) Calculation of discrimination correct rate of discrimination function Next, the process of calculating the discrimination correct rate of discrimination function is as follows. That is, the sound identification condition setting support device 1 includes a plurality of input sampled sounds that are not used for the discrimination function generation process, and a plurality of input soundless sampled sounds. Among them, a non-sound sample sound that is not used in the generation process of the discrimination function is used, the discrimination function is executed to discriminate the presence or absence of the abnormal sound, and the correctness of the discrimination is determined. Specifically, the sound identification condition setting support device 1 includes spectrogram data of the different sampled sounds that are not used in the generation process of the discrimination function among the plurality of input different sampled sounds, and the plurality of input For each spectrogram data of the non-sound sample sound that is not used in the discrimination function generation process, the predetermined frequency range is divided into the number of blocks indicated by the set number of blocks. Of the divided blocks, histogram data indicating a histogram of the sound intensity distribution is generated only for the block in the target frequency band. If the histogram data has already been generated for each block of the spectrogram data of the input abnormal sound sample in order to perform the process of selecting the target frequency band described above, it is determined whether the identification function is correct or not. It is not necessary to generate histogram data anew to perform the process, and from the histogram data generated to select the target frequency band, the spectrogram data of the sampled sound that is not used for the discrimination function generation process What is necessary is just to select and use the histogram data of the target frequency band.

  Subsequently, the sound identification condition setting support device 1 sequentially inputs the frequency of the target intensity in the generated histogram data of the target frequency band to the identification function, and each time it is input, the noise identification is performed by the identification function. The correctness of the identification result is determined. Subsequently, the sound identification condition setting support device 1 calculates an identification accuracy rate that is a probability of correctly identifying presence / absence of an abnormal sound for the identification function based on the obtained determination result.

  The sound identification condition setting support device 1 changes the setting of the number of blocks, the setting of the number of selections c, and the setting of the number of selections d, (1) selection of the target frequency band, (2) selection of the target intensity, (3 ) Generate discriminant function, and (4) calculate the discriminant accuracy rate of discriminant function multiple times. As a result, a plurality of discriminant functions are generated, a discrimination correctness determination is performed for each of the plurality of discriminant functions, and a discrimination correct answer rate is calculated for each of the plurality of discriminant functions from the result.

(5) Selection of deterministic discriminant function etc. Next, the process of selecting the deterministic discriminant function etc. is generally as follows. That is, the sound identification condition setting support device 1 selects the identification function having the highest classification accuracy rate from among the plurality of identification functions generated as described above, the identification function, the number of blocks corresponding to the identification function, The target frequency band corresponding to the discriminant function and the target intensity corresponding to the discriminant function are stored in the storage unit 21 as a definitive discriminant function, the number of definite blocks, a definitive target frequency band, and a definitive target intensity. Thus, the sound identification condition setting support device 1 selects and stores the definite identification function, the number of definite blocks, the definite target frequency band, and the deterministic target intensity. 30 (see FIG. 18).

  Here, the spectrogram of sound and the histogram of intensity distribution will be described with reference to FIGS. FIG. 5 shows a specific example of the waveform of the abnormal sound sample. In FIG. 5, the horizontal axis represents time, and the vertical axis represents sound intensity. FIG. 6 shows a spectrogram of the abnormal sample sound shown in FIG. In FIG. 6, the horizontal axis represents time, the vertical axis represents frequency, and the color represents the intensity of sound. In FIG. 6, the sound intensity is represented in two stages of black and white, and black indicates that the sound intensity is greater than white, but the actual spectrogram graph is The intensity of sound is represented by a difference in color (for example, lightness of color) in multiple levels (for example, 256 levels).

  7 shows a histogram of sound intensity distribution in block F1 in FIG. 6, FIG. 8 shows a histogram of sound intensity distribution in block F2 in FIG. 6, and FIG. 9 shows block F3 in FIG. The histogram of the sound intensity distribution in is shown. In each of FIGS. 7 to 9, the vertical axis represents frequency / total power, and the horizontal axis represents sound intensity (intensity). Here, the noise sample sound and the non-noise sample sound have the same length (time from the start time to the end time), and the length of the sound sample sound or the sound sample sound without noise is the same. A value obtained by dividing (for example, 1 hour) by a predetermined unit time (for example, 1 second) corresponds to “total frequency”. In addition, it is obtained by counting the number of times that the sound intensity is the same in the predetermined unit time from the start point to the end point of the abnormal sound sample or non-noise sample sound. The obtained value corresponds to “frequency”.

  In the spectrogram of the abnormal sound sample shown in FIG. 6, in the block F1, the state where the sound intensity is high is continuous between the start time and the end time of the abnormal sound sample. In the sound intensity distribution in this block F1, as shown in FIG. 7, the sound intensity at which the frequency / total frequency is maximum (peak), that is, the maximum frequency intensity is P1, the distribution is concentrated on P1, and the distribution The range of is narrow.

  Further, in the spectrogram of the abnormal sound sample shown in FIG. 6, in the block F2, the state in which the sound intensity is low is continuous from the start time to the end time of the abnormal sound sample. In the sound intensity distribution in the block F2, as shown in FIG. 8, the maximum frequency intensity is P2, and P2 is a value smaller than P1. Further, in the sound intensity distribution in the block F2, the distribution is concentrated on P2, and the distribution range is narrow.

  Further, in the spectrogram of the abnormal sound sample shown in FIG. 6, in block F3, the sound intensity is high and the sound intensity is low between the start time and the end time of the noise sample sound. Are mixed. In the sound intensity distribution in the block F3, as shown in FIG. 8, the range of the distribution is relatively wide.

  As described above, according to the histogram of the intensity distribution for each block in the spectrogram of the abnormal sound sample, the temporal change in the sound intensity for each block of the abnormal sound sample can be expressed. Therefore, by using the histogram of the sound intensity distribution for each block in the spectrogram for each sampled sound, such as abnormal sound sample, non-natural sound sample sound, to-be-determined sound described later, it is included in the sampled sound and sampled sound. It is possible to generate a discriminant function that can accurately identify the presence or absence of an abnormal sound in a sampled sound while taking into account temporal changes in the frequency or intensity of the abnormal sound.

(Configuration of sound identification condition setting support device)
As shown in FIG. 1, the sound identification condition setting support device 1 includes an arithmetic processing unit 11. The arithmetic processing unit 11 divides the spectrogram data into a plurality of blocks, and a spectrogram generation unit 12 that generates spectrogram data from speech data. A histogram generation unit 13 that generates histogram data of sound intensity distribution in each block, a block selection unit 14 that selects a candidate block from a plurality of blocks of each spectrogram data of the abnormal sound sample sound, and an abnormal sound sample A frequency band selection unit 15 that selects a target frequency band for spectrogram data of sound, an intensity selection unit 16 that selects target intensity for spectrogram data of abnormal sound samples, and a sound intensity distribution in a block of the target frequency band Using the frequency of object intensity in the histogram data of A discriminant function generating unit 17 for generating a discriminant function, a correct rate calculating unit 18 for calculating a discriminant accuracy rate of the discriminant function, a discriminant function having the highest discriminant correct rate, the number of blocks corresponding to the discriminant function, and the discriminant function Identification condition selection unit 19 for selecting a corresponding target frequency band and a target intensity corresponding to the identification function, a process for setting or changing the number of blocks and the number of selections, a flow control of a sound identification condition setting support process, etc. The control part 20 which performs is provided. Furthermore, the sound identification condition setting support device 1 includes a storage unit 21 that stores a definite identification function, the number of definite blocks, a definite target frequency band, a deterministic target strength, and various data to be held during the sound identification condition setting support process; And an input / output unit 22 that performs input of voice data of abnormal sampled sound or non-sound sampled sound, operation input, display of progress of processing, and the like.

  The sound identification condition setting support device 1 having such a configuration can be realized by a personal computer, for example. More specifically, the arithmetic processing unit 11 can be realized by a CPU (Central Processing Unit) built in a personal computer. Further, the spectrogram generation unit 12, the histogram generation unit 13, the block selection unit 14, the frequency band selection unit 15, the intensity selection unit 16, the discrimination function generation unit 17, the accuracy rate calculation unit 18, the discrimination condition selection unit 19, and the control unit 20 Can be realized by causing a CPU to read and execute a computer program for causing a personal computer to function as the sound identification condition setting support device 1. The storage unit 21 can be realized by a storage device (for example, a hard disk drive device, a flash memory, etc.) included in the personal computer. The input / output unit 22 can be realized by an input / output device (for example, a keyboard, a mouse, a monitor, etc.) included in the personal computer.

  The histogram generation unit 13 is a specific example of a setting abnormal sound histogram generation unit, a generation histogram generation unit, and a verification histogram generation unit.

(Sound identification condition setting support processing)
2 to 4 show a specific flow of the sound identification condition setting support process performed by the sound identification condition setting support device 1. The sound identification condition setting support process outlined above will be described more specifically with reference to FIGS. 2 to 4 and the like.

  First, as a preparation for processing, for example, the worker prepares audio data of 100 abnormal sound samples A1 to A100 and audio data of 100 non-natural sound samples B1 to B100.

  Subsequently, the operator operates the input / output unit 22 of the sound identification condition setting support device 1 to obtain the sound data of the abnormal sound samples A1 to A100 and the sound data of the non-natural sound samples B1 to B100 as sound identification conditions. Input to the setting support apparatus 1. When these sound data are input to the sound identification condition setting support device 1, the sound identification condition setting support device 1 receives the sound data of the abnormal sound samples A1 to A100 and the non-natural sound samples B1 to B100. In addition to being able to identify each one, it is possible to identify whether the sound data is the sound data of an abnormal sound sample or the sound data of an allophone sample sound. As a method for realizing this, for example, each voice data is identified with identification information for identifying each voice data, and whether each voice data is voice data of abnormal sample sound or voice data of non-sound sample sound. There is a method of attaching classification information indicating whether or not there is. This method can be easily realized, for example, by devising a method for naming the file name of the audio data.

  Subsequently, the worker inputs an instruction to start the sound identification condition setting support process to the sound identification condition setting support device 1. In response to this, the sound identification condition setting support device 1 starts the sound identification condition setting support process shown in FIGS.

  In the sound identification condition setting support process, as shown in FIG. 2, first, the spectrogram generation unit 12 generates spectrogram data for each of the sound data of the abnormal sound samples A1 to A100 and the non-natural sound samples B1 to B100. Then, the generated spectrogram data is stored in the storage unit 21 (step S1).

  Subsequently, the sound identification condition setting support device 1 performs a process of selecting a target frequency band (steps S2 to S6). That is, first, the control unit 20 sets the number of blocks when the predetermined frequency range of the spectrogram data is divided into a plurality of blocks according to a predetermined rule (step S2). For example, in the present embodiment, as the predetermined rule, the minimum number of blocks is set to 20, the maximum value is set to 80, the number of blocks set for the first time is set to the minimum value, and the setting of the number of blocks is changed once. A rule of increasing the number of blocks by 5 every time is adopted.

  Here, FIG. 10 shows a state in which the number of blocks is set to 20 and a predetermined frequency range of the spectrogram of the collected sound is divided into blocks G, and FIG. 11 shows a state in which the number of blocks is set to 30 and the spectrogram of the collected sound is predetermined. The frequency range is divided into blocks H. As can be seen from FIGS. 10 and 11, as the number of blocks increases, the width (frequency band range) of each block decreases.

  Subsequently, as shown in FIG. 2, the control unit 20 reads spectrogram data of each of the abnormal sound samples A1 to A100 from the storage unit 21. Then, the histogram generation unit 13 divides the predetermined frequency range of the spectrogram data of each of the abnormal sound samples A1 to A100 that has been read out into the number of blocks indicated by the number of blocks set in step S2. The generated histogram data of the intensity distribution of each block is generated, and the generated histogram data is stored in the storage unit 21 (step S3). Thereby, histogram data of intensity distribution is generated and stored for each block for the spectrogram data of the abnormal sound samples A1 to A100.

  Subsequently, the block selection unit 14 reads out the histogram data of the spectrogram data of each abnormal sound sample A1 to A100 from the storage unit 21, and in each read spectrogram data, all except the block having the lowest frequency band. For each block, the maximum frequency strength of the histogram data of two adjacent blocks is compared, and the maximum frequency strength of the histogram data of the block with the higher frequency band is the maximum frequency strength of the histogram data of the block with the lower frequency band If it is greater than the intensity, the block with the higher frequency band is selected as a candidate block (step S4). Thereby, a candidate block is selected for each of the spectrogram data of the abnormal sound samples A1 to A100.

Here, FIG. 12 shows the frequency band, the maximum frequency intensity of the histogram data, and the flag indicating that the candidate block is selected for each block of the spectrogram data of the abnormal sound sample A1. For convenience of explanation, block numbers # 1 to #n are assigned to the respective blocks. In FIG. 12, the predetermined frequency range of the spectrogram data is α kHz to β kHz, and the number of blocks is n. For example, the maximum power intensity of the block # 2 is 1 × 10 ^−12, which is higher than the maximum power intensity 1 × 10 ^{−13 of the} adjacent block # 1 whose frequency band is lower than that of the block # 2. Therefore, block # 2 is selected as a candidate block. Similarly, blocks # 3, # 6, and # 7 are also selected as candidate blocks.

  Subsequently, as shown in FIG. 2, the control unit 20 sets the selection number c relating to the selection of the frequency band according to a predetermined rule (step S5). For example, in the present embodiment, as the predetermined rule, the minimum value of the selection number c is 3, the maximum value is 15, the selection number c set for the first time is the minimum value, and the setting of the selection number c is changed. A rule is adopted in which the number of selections c is increased by 1 each time it is performed.

  Subsequently, the frequency band selection unit 15 counts the number of candidate blocks in the same frequency band in the spectrogram data of the abnormal sound samples A1 to A100, and the set selection number c indicates in descending order of the count value. The number of frequency bands are specified, and the specified frequency band and the adjacent frequency band lower than the specified frequency band are selected as target frequency bands (step S6).

  Here, FIG. 13 shows the selection status of candidate blocks in the spectrogram data of abnormal sound samples A1 to A100 and the count values of candidate blocks in the same frequency band. Similarly to FIG. 12, for convenience of explanation, block numbers # 1 to #n are assigned to the respective blocks. In FIG. 13, the frequency band of block # 3 is selected as the target frequency band because the count value is 30 and the count value is the largest. Further, the frequency band of the adjacent block # 2 whose frequency band is lower than that of the block # 3 is also selected as the target frequency band. The frequency band of block # 7 is selected as the target frequency band because the count value is 28 and the count value is the second largest. Further, the frequency band of the adjacent block # 6 whose frequency band is lower than that of the block # 7 is also selected as the target frequency band.

  Subsequently, the sound identification condition setting support device 1 performs a process of selecting a target intensity (steps S7 to S9 in FIG. 2). That is, first, the control unit 20 sets the selection number d related to the selection of intensity according to a predetermined rule (step S7). For example, in the present embodiment, as the predetermined rule, the minimum value of the selection number d is set to 3, the maximum value is set to 8, the selection number d set for the first time is set to the minimum value, and the setting of the selection number d is changed. A rule is adopted in which the number of selections d is increased by 1 each time it is performed.

  Subsequently, the control unit 20 reads spectrogram data of each of the unusual sound samples A1 to A100 from the storage unit 21. Then, the intensity selection unit 16 specifies the histogram data of the block of the target frequency band selected in step S6 from the histogram data in each read spectrogram data, and the frequency of the intensity of each sound in the histogram data Is ranked (step S8). That is, in one histogram data, ranking is performed for each histogram data of the target frequency band such that the highest frequency intensity is first and the next highest intensity is second.

  Subsequently, the intensity selection unit 16 calculates the average rank (average rank) of the frequency of sound intensity in the histogram data of the same target frequency band in the spectrogram data of the abnormal sound samples A1 to A100. In the descending order, the number of intensities indicated by the selection number d set in step S7 is specified, and these specified intensities are selected as target intensities (step S9). When a plurality of target frequency bands are selected, the calculation of the average rank and the selection of the target strength in step S9 are performed for each target frequency band.

Here, FIG. 14 shows the respective intensities of one histogram data among the spectrogram data of the abnormal sound samples A1 to A100, and the rankings assigned to these intensities. In FIG. 14, for the convenience of explanation, the strength numbers # 1 to #m are assigned to the strengths. According to FIG. 14, the intensity of # 7 is first and the intensity of # 6 is second. As can be seen from FIG. 14, the “intensity” selected by the intensity selection unit 16 is strictly an intensity range having a certain width. For example, the intensity of # 1 has a width of 1 × 10 ⁻²⁰ to 1 × 10 ⁻¹⁹ .

  FIG. 15 shows the average rank of the sound intensity frequencies in the histogram data of the blocks of the same target frequency band in the spectrogram data of the abnormal sound samples A1 to A100. In FIG. 15, the average rank of intensity of # 7 is 1.3, and the average rank of intensity of # 6 is 2.2.

  Subsequently, the sound identification condition setting support device 1 performs a process of generating an identification function (steps S10 to S12). That is, first, as shown in FIG. 2, the control unit 20 randomly selects 50 spectrogram data of sampled abnormal sounds, and the histogram data of the block of the target frequency band in each selected spectrogram data from the storage unit 21. Read (step S10). Subsequently, the control unit 20 stores, in the storage unit 21, history information indicating which abnormal sound sample data is selected from the spectrogram data of the abnormal sound samples A1 to A100. This history information is used by the control unit 20 to specify spectrogram data that has not been selected in step S13.

  Subsequently, as shown in FIG. 3, the control unit 20 randomly selects 50 spectrogram data of the allophone sample sound, and reads each selected spectrogram data from the storage unit 21. Then, the histogram generation unit 13 divides the predetermined frequency range of each read spectrogram data into the number of blocks indicated by the number of blocks set in step S2, and among the divided blocks, the target frequency band Histogram data of the intensity distribution of the block is generated, and each generated histogram data is stored in the storage unit 21 (step S11). As a result, histogram data of the intensity distribution of the block in the target frequency band is generated and stored for each of the spectrogram data of 50 randomly selected sample sounds. Subsequently, the control unit 20 stores, in the storage unit 21, history information indicating which of the non-sound sample sounds the spectrogram data has been selected from the spectrogram data of the non-sound sample sounds B <b> 1 to B <b> 100. This history information is used by the control unit 20 to specify spectrogram data that has not been selected in step S17.

  Subsequently, the discriminant function generation unit 17 performs histogram data of the target frequency band in the spectrogram data of the 50 abnormal sound samples read out in step S10 and the 50 non-noise samples generated in step S11. Using the histogram data of the target frequency band in the sound spectrogram data, a discrimination function is generated. At this time, the discriminant function generation unit 17 generates the discriminant function using only the frequency of the target intensity in each histogram data (step S12). As described above, in step S12, 50 spectrogram data randomly selected from the spectrogram data of the abnormal sound samples A1 to 100 and the spectrogram data of the non-sound sample sounds B1 to B100 are selected at random. The discriminant function is generated from the 50 spectrogram data. At this time, the histogram data used for generating the discriminant function is only the frequency of the target intensity in the histogram data of the block in the target frequency band.

  In addition, for generating the discriminant function, for example, a pattern recognition method using a learning model such as a support vector machine (see, for example, JP-A-2005-352997) is adopted. Note that another learning model such as a neural network may be employed for generating the discrimination function.

  Subsequently, the identification function generation unit 17 generates the generated identification function, the number of blocks corresponding to the identification function (for example, information such as a numerical value indicating the number of blocks set when the identification function is generated), and the identification The selection number c corresponding to the function (for example, information such as a numerical value indicating the selection number c set when the identification function is generated) and the selection number d corresponding to the identification function (for example, when the identification function is generated) Information such as a numerical value indicating the set selection number d), a target frequency band corresponding to the identification function (for example, information indicating a target frequency band selected when the identification function is generated), and the identification function Is stored in the storage unit 21 (for example, information indicating the target intensity selected when the identification function is generated).

  Subsequently, the sound identification condition setting support device 1 performs processing for calculating the identification accuracy rate of the identification function (steps S13 to S22). That is, as shown in FIG. 3, first, the control unit 20 randomly selects spectrogram data of unselected abnormal sound samples, that is, spectrogram data of abnormal sound samples not selected in step S10. In the selected spectrogram data, the histogram data of the block in the target frequency band is read from the storage unit 21 (step S13).

  Subsequently, the accuracy rate calculation unit 18 inputs the frequency of the target intensity in the histogram data of the target frequency band read out in step S13 to the discrimination function generated in step S12, and the abnormal sound sample is input by the discrimination function. The presence / absence of abnormal noise in the sound is identified (step S14).

  Subsequently, the correct rate calculation unit 18 determines whether or not the identification function is correct (step S15). In the determination in step S15, the answer is correct when the result of the identification by the identification function is “abnormal noise”, and the answer is incorrect otherwise. Subsequently, the correct rate calculation unit 18 stores the determination result (for example, information such as a flag indicating the determination result) in the storage unit 21 as determination result information.

  Subsequently, the control unit 20 uses the spectrogram data of the unselected phantom sample sound among the spectrogram data of the phantom sample sounds A1 to A100, that is, the phantom sample that has not been selected in step S10 or step S13. It is determined whether or not there is sound spectrogram data (step S16). Then, when there is spectrogram data of the unselected abnormal sound sample (step S16: YES), the control unit 20 returns the process to step S13. On the other hand, when there is no spectrogram data of the unselected abnormal sound sample (step S16: NO), the control unit 20 shifts the process to step S17. As described above, by steps S13 to S16, among the spectrogram data of the abnormal sound samples A1 to A100, the spectrogram data of the abnormal sound samples that were not used for generating the discriminant function is used, and whether or not the discrimination by the discriminant function is correct. Determined. At this time, the histogram data used for the determination is only the frequency of the target intensity in the histogram data of the target frequency band.

  Subsequently, the control unit 20 randomly selects one of the spectrogram data of the unselected sound sample sound, that is, the spectrogram data of the sound sample sound not selected in step S11, and stores the selected spectrogram data. Read from the unit 21. Then, the histogram generation unit 13 divides the predetermined frequency range of the read spectrogram data into the number of blocks indicated by the number of blocks set in step S2, and among the divided blocks, blocks in the target frequency band The histogram data of the intensity distribution is generated, and each generated histogram data is stored in the storage unit 21 (step S17).

  Subsequently, the correct answer rate calculation unit 18 inputs the frequency of the target intensity in the histogram data of the block of the target frequency band generated in step S17 to the discrimination function generated in step S12, and the allophone is generated by the discrimination function. The presence / absence of abnormal noise in the sample sound is identified (step S18).

  Subsequently, the correct answer rate calculation unit 18 determines whether the identification function is correct or not (step S19). In the determination in step S <b> 16, the answer is correct when the result of the identification by the identification function is “no abnormal sound”, otherwise it is incorrect. Subsequently, the correct answer rate calculation unit 18 stores the determination result in the storage unit 21 as determination result information.

  Subsequently, the control unit 20 sets the spectrogram data of the unselected non-sound sample sound among the spectrogram data of the non-sound sample sounds B1 to B100, that is, the non-sound sample sound that is not selected in step S11 or step S17. It is determined whether there is any spectrogram data (step S20). If there is spectrogram data of an unselected sound sample sound that has not been selected (step S20: YES), the control unit 20 returns the process to step S17. On the other hand, when there is no spectrogram data of the unselected sound sample sound that has not been selected (step S20: NO), the control unit 20 shifts the process to step S21. As described above, by steps S17 to S20, the spectrogram data of the non-sound sample sound that was not used for the generation of the discrimination function among the spectrogram data of the non-sound sample sounds B1 to B100 is used to determine whether the discrimination function is correct or not. Determined. At this time, the histogram data used for the determination is only the frequency of the target intensity in the histogram data of the target frequency band.

  Subsequently, in step S21 in FIG. 4, the control unit 20 determines whether or not the discrimination function generation and the discrimination correctness determination by the discrimination function (steps S10 to S20) have been repeated a predetermined number of times, for example, 200 times. If the generation of the discriminant function and the determination of correctness of discrimination by the discriminant function have not been repeated 200 times (step S21: NO), the control unit 20 returns the process to step S10 to redo the random selection of spectrogram data from the beginning. All the history information indicating the history of selection of spectrogram data is erased (initialized). On the other hand, when the generation of the discriminant function and the determination of correctness of discrimination by the discriminant function are repeated 200 times (step S21: YES), the control unit 20 shifts the process to step S22.

  Through the processes in steps S10 to S21 described above, there are 200 combinations of 50 different sample sounds used for generating the discrimination function and 50 combinations of 50 non-sampled sounds used for generating the discrimination function. Thus, 200 discriminant functions are generated using these 200 combinations. Further, by the processing of steps S10 to S21, a combination of 50 different sampled sounds used for determination of correctness of discrimination by the discriminant function and 50 non-sound sample sounds used for determination of correctness of discrimination by the discriminant function. 200 combinations are generated. Using these 200 combinations, it is determined whether each identification function is correct or not. When Steps S10 to S21 have been executed 200 times, the storage unit 21 stores 20000 pieces of determination result information. In this way, by repeatedly generating the discrimination function and determining whether the discrimination function is correct or not, and obtaining a large number of pieces of determination result information, the reliability of the discrimination correct answer rate calculated later in step S22 can be improved.

  Subsequently, in step S22, the correct answer rate calculation unit 18 uses the 20000 pieces of determination result information obtained in steps S10 to S21 to set one block number, one selection number c, and one selection number d. The discrimination accuracy rate is calculated for the discrimination function generated under the current state, and the calculated discrimination accuracy rate is stored in the storage unit 21.

  Here, FIG. 16 shows an example of the method of calculating the identification correct rate in step S22. In FIG. 16, “◯” indicates that the determination result in step S15 or S19 is correct, and “X” indicates that the determination result in step S15 or S19 is incorrect. “-” Indicates that the spectrogram data of the sample sound is used for generating the discriminant function, and is not used for the discrimination correctness judgment by the discriminant function in step S15 or S19.

  As shown in FIG. 16, in the discrimination correct rate calculation process in step S22, first, the discrimination correct rate is calculated for the spectrogram data of the sample sounds A1 to A100 and B1 to B100. The classification accuracy rate for one spectrogram data is the sum of the number of correct answers and the number of incorrect answers in the identification result obtained by inputting the histogram data of the spectrogram data into the identification function. It can be determined by dividing. Next, an average value of the identification accuracy rates of the spectrogram data of the sample sounds A1 to A100 and B1 to B100 is calculated, and one block number, one selection number c, and one selection number d are set as the average value. It is used as the discrimination accuracy rate of the discriminant function generated under the current state.

  Subsequently, as illustrated in FIG. 4, the control unit 20 determines whether or not the processing of steps S <b> 7 to S <b> 22 has been completed for all settings of the selection number d (step S <b> 23). When the execution of the processing of steps S7 to S22 has not been completed for all the settings of the selection number d (step S23: NO), the control unit 20 returns the processing to step 7 in FIG. In step S7, the control unit 20 changes the selection number d according to the predetermined rule. Subsequently, the processes of steps S8 to S22 are executed. On the other hand, when the execution of the processing of steps S7 to S22 is completed for all the settings of the selection number d (step S23: YES), the control unit 20 shifts the processing to step S24.

  According to the predetermined rule for setting the selection number d described above, since the selection number d is set in six ways in total, the processes of steps S7 to S22 are repeated six times while the selection number d is changed. As a result, the discrimination correct answer rate of the discrimination function generated by the six selection numbers d is calculated.

  Subsequently, the control unit 20 determines whether or not the processing of steps S5 to S23 has been completed for all settings of the selection number c (step S24). If the execution of the processing of steps S5 to S23 has not been completed for all the settings of the selection number c (step S24: NO), the control unit 20 returns the processing to step 5 in FIG. In step S5, the control unit 20 changes the selection number c according to the predetermined rule. Following this, the processes of steps S6 to S23 are executed. On the other hand, when the execution of the processes of steps S5 to S23 is finished for all the settings of the selection number c (step S24: YES), the control unit 20 shifts the process to step S25.

  According to the predetermined rule for setting the selection number c described above, since the selection number c is set in 13 ways in total, the processes of steps S5 to S23 are repeated 13 times while the selection number c is changed. As a result, the discrimination accuracy rate of the discrimination function generated by the 13 selection numbers c is calculated.

  Subsequently, the control unit 20 determines whether or not the processing of steps S2 to S24 has been completed for all settings of the number of blocks (step S25). If the execution of the processes of steps S2 to S24 has not been completed for all settings of the number of blocks (step S25: NO), the control unit 20 returns the process to step S2 in FIG. In step S2, the control unit 20 changes the number of blocks according to the predetermined rule. Subsequently, the processes of steps S3 to S24 are executed. On the other hand, when the execution of the processes of steps S2 to S24 is completed for all settings of the number of blocks (step S25: YES), the control unit 20 shifts the process to step S26.

  According to the predetermined rule for setting the number of blocks described above, there are 13 types of setting of the number of blocks. Therefore, the processes of steps S2 to S24 are repeated 13 times while the number of blocks is changed. Further, in each loop of steps S2 to S24, the processing of steps S5 to S23 is repeated 13 times, and further, in each loop of steps S5 to S23, the processing of steps S7 to S22 is repeated 6 times. The discrimination accuracy rate of the discrimination function generated under the combination of the 13 block numbers, the 13 selection numbers c, and the 6 selection numbers d is calculated. When the execution of the processing of steps S2 to S25 is finished for the setting of 13 types of blocks, the storage unit 21 stores combinations of 13 types of blocks, 13 types of selections c, and 6 types of selections d, respectively. Corresponding 1014 identification correct answer rates are stored.

  Here, FIG. 17 shows an example of the identification correct answer rate of the identification function corresponding to the combination of the number of blocks, the number of selections c, and the number of selections d set according to the predetermined rule.

  Subsequently, the sound identification condition setting support device 1 performs a process of selecting a definite identification function or the like (steps S26 to S27). That is, in step S26 in FIG. 4, the identification condition selection unit 19 reads a plurality of identification accuracy rates stored in the storage unit 21 and identifies the highest identification accuracy rate from among them. Subsequently, the identification condition selection unit 19 selects the number of blocks, the target frequency band, and the target intensity corresponding to the identification function having the identified identification accuracy rate, and determines the selected block number, the target frequency band, and the target intensity as the number of determined blocks. Then, it is stored in the storage unit 21 as the confirmation target frequency band and the confirmation target intensity.

  Subsequently, all the spectrogram data of the abnormal sound samples A1 to A100 and the non-natural sound samples B1 to B100 generated under the setting of the number of blocks selected as the determined number of blocks by the discriminant function generation unit 17. In the histogram data of the frequency band to be determined at, an identification function is generated again using the frequency of the intensity of the determination target, and the generated identification function is stored in the storage unit 21 as the determined identification function (step S27). The number of determined blocks, the frequency band to be determined, the strength to be determined, and the determination function are used for abnormal sound determination processing by the abnormal sound determination device 30 described later.

  As described above, according to the sound identification condition setting support device 1 according to the embodiment of the present invention, an identification function that can accurately identify the presence or absence of an abnormal sound in a collected sound can be quickly and easily performed with a small work load. Can be generated. That is, since the generation of the discrimination function by the sound discrimination condition setting support device 1 is automatically performed as long as voice data is input, the task of generating the discrimination function, that is, the task corresponding to the feature extraction of abnormal noise, Compared with the conventional feature extraction operation of finding the features of abnormal sounds while looking at the spectrogram of the collected sound displayed on the computer monitor, it can be greatly simplified. Therefore, the working time can be shortened and the burden on the worker can be reduced. Also, if a sample sound is prepared, the sound identification condition setting support device 1 can automatically generate a discrimination function, so that a discrimination function having a high accuracy discrimination ability can be obtained regardless of the experience of the operator. Can be generated.

  In the sound identification condition setting support processing according to the embodiment of the present invention, the target frequency band and the target intensity are selected in the spectrogram data of the abnormal sound sample, and the subsequent processing, that is, the spectrogram of each non-noise sample sound. The process of generating the histogram data of the data, the process of generating the discrimination function, and the process of determining whether the discrimination by the discrimination function is correct or not are performed using only the frequency of the target intensity in the histogram data of the block in the target frequency band. Therefore, compared to the case where these processes are performed using all blocks or the case where these processes are performed using all the frequencies of histogram data, the processing amount and the amount of data handled in the processing are greatly increased. Can be reduced. Therefore, according to the sound identification condition setting support processing according to the embodiment of the present invention, a general personal computer is used to generate and select an identification function that can accurately identify the presence or absence of abnormal sounds in the collected sound. Can be done quickly.

  In the sound identification condition setting support processing according to the embodiment of the present invention, the number of divisions (number of blocks) of a predetermined frequency range in the spectrogram data is changed, and the number of frequency bands to be selected as the target frequency band (number of selections c). And a plurality of discriminant functions are generated while changing the number of intensities to be selected as the target intensities (selection number d), and the discriminant function having the highest discrimination accuracy rate is selected from these discriminant functions. Therefore, it is possible to select a discriminant function having a high-accuracy discriminating capability according to the characteristics of abnormal sounds.

  In the sound identification condition setting support process by the sound identification condition setting support device 1 described above, the number of blocks when the predetermined frequency range of the spectrogram data is divided into blocks and the number of frequency bands to be selected as the target frequency band are calculated. An example has been given of the case where a plurality of discriminant functions are generated while changing the selection number c to be determined and the selection number d to determine the number of sound intensities to be selected as the target intensity, but the present invention is not limited to this. . In the sound identification condition setting support process, after the number of blocks is set at the start of the process, the number of blocks may not be changed. Further, after the selection number c is set at the start of the process in the sound identification condition setting support process, the selection number c may not be changed. In the sound identification condition setting support process, the target frequency band may not be selected.

  Further, in the sound identification condition setting support processing by the sound identification condition setting support device 1 described above, in the spectrogram data of the abnormal sound sample sound, the average value of the ranks of the respective sound intensities in the histogram data of the same target frequency band is obtained. As an example, the calculation is performed and the sound intensity is selected as the target intensity in descending order of the average value (step S9 in FIG. 2). However, the present invention is not limited to this. In the spectrogram data of the sampled noise, calculate the cumulative value of each sound intensity rank in the histogram data of the same target frequency band, and select the sound intensity as the target intensity in descending order of the accumulated value. May be.

  Further, in the sound identification condition setting support process in the above-described embodiment, the case where the generation of the identification function and the determination of whether the identification function is correct (steps S10 to S20) is repeated 200 times, for example. Not limited to this, the generation of the discrimination function and the determination of whether the discrimination function is correct or not may be performed once to 199 times, or 201 times or more.

  Further, in the sound identification condition setting support process in the above-described embodiment, after selecting the number of confirmed blocks, the frequency band to be confirmed and the strength to be confirmed, the sound identification condition setting support process is generated under the setting of the number of blocks selected as the number of confirmed blocks. A discrimination function is generated anew using the frequency of the determination target intensity in the histogram data of the determination target frequency band in all the spectrogram data of the abnormal sound samples A1 to A100 and the non-sound sample sounds B1 to B100, and the generated identification The case where a function is selected as a deterministic discriminating function has been described as an example. However, the present invention is not limited to this, and the discriminating function generated in step S12 in FIG. 3 under the setting of the number of blocks selected as the deterministic block number. May be selected as the deterministic identification function.

  In the above-described embodiment, all the spectrogram data of the abnormal sound samples A1 to A100 are used as the abnormal sound spectrogram data for setting for selecting the target frequency band and the target intensity, and the discrimination function is generated. The half of the spectrogram data of the abnormal sound samples A1 to A100 is used as the spectrogram data for generating the sound, and the spectrogram data of the abnormal sound samples A1 to A100 is used as the verification spectrogram data for calculating the discrimination accuracy rate. As an example, the remaining half of the above is used. However, how to distribute the spectrogram data of a plurality of abnormal sound samples prepared for distribution of setting abnormal sound spectrogram data, generation spectrogram data, and verification spectrogram data can be freely determined as appropriate. be able to. For example, half of the spectrogram data of 100 different sample sounds may be used as the setting specific sound spectrogram data. Also, spectrogram data of 150 different sound samples is prepared, and 50 of them are used as setting different sound spectrograms, the other 50 are used as generation spectrogram data, and the remaining 50 are used. It may be used as spectrogram data for verification. Also, spectrogram data of 180 abnormal sound samples is prepared, and the spectrogram data of 120 abnormal sound samples is used as the setting abnormal sound spectrogram data and generation spectrogram data, and the remaining 60 sound samples are generated. The spectrogram data of the different sampled sounds may be used as the spectrogram data for verification. Even when the spectrogram data of the allophone sample sound is used as the spectrogram data for generation or the spectrogram data for verification, the distribution ratio can be freely determined as appropriate.

(Configuration of abnormal sound determination device)
FIG. 18 shows the presence / absence of abnormal noise in the collected sound by using the confirmed identification function, the number of confirmed blocks, the confirmed target frequency band, and the confirmed target intensity generated by the sound identification condition setting support device 1 according to the embodiment of the present invention. The abnormal sound determination apparatus which determines is shown. In FIG. 18, the abnormal sound determination device 30 includes an arithmetic processing unit 40. The arithmetic processing unit 40 divides the spectrogram data into a plurality of blocks by dividing the spectrogram data into a plurality of blocks by generating spectrogram data from the speech data. Among these blocks, the histogram generating unit 42 that generates histogram data of the sound intensity distribution for the block in the determination target frequency band, and the frequency of the determination target intensity in the histogram data of the block in the determination target frequency band are input to the identification function. , An abnormal sound identification unit 43 for identifying presence / absence of abnormal noise, and a control unit 44 for controlling the flow of abnormal noise determination processing and the like. Further, the abnormal sound determination device 30 has a storage unit 45 for storing various data to be held during the abnormal sound determination processing, and input / output for performing input of sound data of the sound to be determined, operation input, display of processing results, and the like. Part 46.

  The abnormal sound determination device 30 can be realized by a personal computer, for example. Specifically, the arithmetic processing unit 40 can be realized by a CPU built in a personal computer. The spectrogram generation unit 41, the histogram generation unit 42, the abnormal sound identification unit 43, and the control unit 44 can be realized by causing a CPU to read and execute a computer program. The storage unit 45 can be realized by a storage device such as a hard disk drive device of a personal computer. The input / output unit 46 can be realized by an input / output device such as a keyboard or a monitor included in the personal computer. Note that the sound identification condition setting support device 1 and the abnormal sound determination device 30 may be simultaneously operated by one personal computer.

(Abnormal sound judgment processing)
FIG. 19 shows an abnormal sound determination process performed by the abnormal sound determination device 30. The abnormal sound determination device 30 can determine the presence or absence of an abnormal sound for a collected sound over a relatively long time (for example, 24 hours) collected in a place close to a power transmission line or a device, for example. The entire collected sound that is the target of the abnormal sound determination is referred to as a sound to be determined. In the abnormal sound determination processing by the abnormal sound determination device 30, the sound data of the sound to be determined is divided into several sections of a relatively short time (for example, 1 hour), and the presence or absence of the abnormal sound is identified for each section.

  As preparation for the abnormal sound determination process, the operator inputs the confirmed identification function, the number of confirmed blocks, the determined frequency band, and the determined target intensity generated by the sound identification condition setting support device 1 to the abnormal sound determination device 30. These are stored in the storage unit 45 of the abnormal sound determination device 30. Further, the worker inputs the sound data of the sound to be determined to the abnormal sound determination device 30 and stores it in the storage unit 45 of the abnormal sound determination device 30. Subsequently, when the operator inputs an instruction to start the abnormal sound determination process to the abnormal sound determination apparatus 30, the abnormal sound determination apparatus 30 starts the abnormal sound determination process.

  As shown in FIG. 19, in the abnormal sound determination process, first, the control unit 44 reads a definite identification function, the number of definite blocks, a definite target frequency band, and a definite target intensity from the storage unit 45 (step S41).

  Subsequently, the control unit 44, in the sound data of the sound to be determined, among the sections that have not yet been identified for the presence or absence of the abnormal sound (the section closest to the start time of the sound data of the sound to be determined) Is extracted (step S42).

  Subsequently, the spectrogram generation unit 41 generates spectrogram data of the voice data extracted in step S42 (step S43).

  Subsequently, for the spectrogram data generated in step S43, the histogram generation unit 42 divides a predetermined frequency range into blocks of the determined number of blocks, and among each of the divided blocks, for each block of the determined target frequency band, Histogram data of the sound intensity distribution is generated (step S44).

  Subsequently, the abnormal sound identification unit 43 inputs the frequency of the determination target intensity in the histogram data of each block of the determination target frequency to the determination identification function, identifies the presence or absence of abnormal noise, and stores the identification result in the storage unit 45. (Step S45).

  Subsequently, the control unit 44 determines whether or not the identification of the presence or absence of abnormal sounds has been completed for all sections of the sound data of the sound to be determined (step S46). If the identification of the presence / absence of abnormal sound has not been performed for all sections of the sound data of the sound to be determined (step S46: NO), the process returns to step S42, and the control unit 44 performs the sound corresponding to the next section. Data is extracted from the sound data of the sound to be determined, and the spectrogram generation unit 41, the histogram generation unit 42, and the abnormal sound identification unit 43 execute the processes of steps S43 to S45.

  On the other hand, when the identification of the presence / absence of abnormal sound has been completed for all sections of the sound data of the sound to be determined (step S46: YES), the control unit 44 identifies the presence / absence of abnormal sound for the sound to be determined. The result is displayed on a monitor or the like for each section (step S47). By looking at the monitor, the operator can immediately grasp the presence or absence of abnormal noise in the sound to be determined. In addition, since the presence / absence of abnormal noise in the judged sound can be grasped for each section, the collected sound used as the judged sound is relatively long, and the abnormal sound included therein is relatively short. Even in this case, it is possible to quickly find an abnormal sound in the collected sound.

  In addition, according to the abnormal sound determination device 30, in the spectrogram data of the sound to be determined, histogram data is generated only for the block in the fixed frequency band, and the determined target intensity in the histogram data of the block in the fixed frequency band is determined. Since discrimination using the discrimination function is performed only for the frequency, when generating histogram data for all blocks of the spectrogram data of the sound to be judged, or using the discrimination function for all intensities of the histogram data in the spectrogram data of the sound to be judged Compared with the case of executing the above, it is possible to greatly speed up the processing.

  It should be noted that the present invention can be appropriately changed without departing from the spirit or idea of the invention that can be read from the claims and the entire specification, and the sound identification condition setting support device and the sound identification condition accompanying such a change. A setting support method is also included in the technical idea of the present invention.

DESCRIPTION OF SYMBOLS 1 Sound identification condition setting assistance apparatus 11 Arithmetic processing part 12 Spectrogram generation part 13 Histogram generation part 14 Block selection part 15 Frequency band selection part 16 Intensity selection part 17 Discrimination function generation part 18 Correct answer rate calculation part 19 Identification condition selection part 20 Control part DESCRIPTION OF SYMBOLS 21 Memory | storage part 22 Input / output part 30 Abnormal sound determination apparatus 40 Arithmetic processing part 41 Spectrogram generation part 42 Histogram generation part 43 Abnormal sound identification part 44 Control part 45 Storage part 46 Input / output part

Claims (5)

A sound identification condition setting support device that supports setting of an identification condition for identifying the presence or absence of abnormal noise,
It receives setting noise spectrogram data showing the spectrograms of multiple setting abnormal sound samples including abnormal sounds, and divides the predetermined frequency range into multiple blocks in each setting abnormal sound spectrogram data. A setting noise histogram generator for generating setting noise histogram data indicating a histogram of the sound intensity distribution in each block;
In the histogram data of the blocks in the same frequency band in the plurality of setting abnormal sound spectrogram data in which the setting abnormal sound histogram data is generated by the setting abnormal sound histogram generation unit, the frequency of the intensity of each sound An intensity selection unit that calculates an average value or cumulative value of ranks, specifies a predetermined selection number of sound in descending order of the average value or cumulative value, and selects the specified sound intensity as a target intensity; ,
Receiving generation spectrogram data indicating spectrograms of a plurality of generation sample sounds including generation abnormal sound samples including abnormal sounds and generation non-abnormal sound samples including no abnormal sounds, and generating each spectrogram data A generation histogram generation unit that divides a predetermined frequency range into a plurality of blocks and generates generation histogram data indicating a histogram of the sound intensity distribution in each block;
An identification function generator for generating an identification function for identifying the presence or absence of abnormal noise using the frequency of the target intensity in each of the generation histogram data;
It receives spectrogram data for verification indicating spectrograms of a plurality of sample sounds for verification including abnormal sound samples for verification including abnormal sounds and non-noise sample sounds for verification not including abnormal sounds, and each of the spectrogram data for verification A verification histogram generation unit that divides a predetermined frequency range into a plurality of blocks and generates verification histogram data indicating a histogram of sound intensity distribution in each block;
The frequency of the target intensity in each verification histogram data is input to the discrimination function to identify the presence / absence of abnormal noise, and the discrimination accuracy rate that is the probability that the discrimination function correctly identifies the presence / absence of abnormal noise is calculated. A sound identification condition setting support device, comprising:   While changing the setting of the number of selections, the processing by the intensity selection unit, the generation histogram generation unit, the discriminant function generation unit, the verification histogram generation unit, and the accuracy rate calculation unit is executed a plurality of times. As a result of executing the identification function a plurality of times, the identification function having the highest identification accuracy rate among the plurality of identification functions generated by the identification function generation unit and the target strength selected when generating the identification function are determined as the identification condition. The sound identification condition setting support apparatus according to claim 1, further comprising: an identification condition selection unit that selects the sound identification condition as a sound identification condition selection unit.   While changing the setting of the number of blocks and the selection number when dividing a predetermined frequency range in each spectrogram data into a plurality of blocks, the setting unusual sound histogram generation unit, the intensity selection unit, and the generation histogram A plurality of identifications generated by the discrimination function generation unit as a result of executing the processing by the generation unit, the discrimination function generation unit, the verification histogram generation unit, and the accuracy rate calculation unit a plurality of times and executing the processing a plurality of times. The discrimination function having the highest discrimination accuracy rate among the functions, the number of blocks set when generating the discrimination function, and the target strength selected when generating the discrimination function are selected as the discrimination conditions. The sound identification condition setting support device according to claim 1, further comprising an identification condition selection unit. A sound identification condition setting support method for supporting setting of an identification condition for identifying the presence or absence of abnormal noise,
It receives setting noise spectrogram data showing the spectrograms of multiple setting abnormal sound samples including abnormal sounds, and divides the predetermined frequency range into multiple blocks in each setting abnormal sound spectrogram data. A setting noise histogram generation step for generating setting noise histogram data indicating a histogram of the sound intensity distribution in each block;
In the histogram data of the blocks of the same frequency band in the plurality of setting noise spectrogram data for which the setting noise histogram data has been generated in the setting noise histogram generation step, the frequency of the intensity of each sound A strength selection step of calculating an average value or cumulative value of ranks, specifying a predetermined number of sound intensities in descending order of the average value or cumulative value, and selecting the identified sound strengths as target strengths; ,
Receiving generation spectrogram data indicating spectrograms of a plurality of generation sample sounds including generation abnormal sound samples including abnormal sounds and generation non-abnormal sound samples including no abnormal sounds, and generating each spectrogram data A generation histogram generation step of dividing a predetermined frequency range into a plurality of blocks and generating generation histogram data indicating a histogram of the sound intensity distribution in each block;
An identification function generating step for generating an identification function for identifying the presence or absence of abnormal noise using the frequency of the target intensity in each of the generation histogram data;
It receives spectrogram data for verification indicating spectrograms of a plurality of sample sounds for verification including abnormal sound samples for verification including abnormal sounds and non-noise sample sounds for verification not including abnormal sounds, and each of the spectrogram data for verification A verification histogram generating step of dividing a predetermined frequency range into a plurality of blocks and generating verification histogram data indicating a histogram of the sound intensity distribution in each block;
The frequency of the target intensity in each verification histogram data is input to the discrimination function to identify the presence / absence of abnormal noise, and the discrimination accuracy rate that is the probability that the discrimination function correctly identifies the presence / absence of abnormal noise is calculated. A correct answer rate calculation process,
While changing the setting of the selection number, the processing by the intensity selection step, the generation histogram generation step, the discriminant function generation step, the verification histogram generation step and the accuracy rate calculation step is executed a plurality of times. As a result of executing a plurality of times, the discriminant function having the highest discrimination accuracy rate among the plurality of discriminant functions generated in the discriminant function generation step and the target strength selected when generating the discriminant function are set as the discriminant condition. Sound identification condition setting support method characterized by selecting as   While changing the setting of the number of blocks and the number of selections when dividing a predetermined frequency range in each spectrogram data into a plurality of blocks, the setting abnormal sound histogram generation step, the intensity selection step, and the generation histogram The identification function generation step, the verification histogram generation step, the verification histogram generation step, and the accuracy rate calculation step are executed a plurality of times. As a result of executing the processing a plurality of times, a plurality of identifications generated in the discrimination function generation step The discrimination function having the highest discrimination accuracy rate among the functions, the number of blocks set when generating the discrimination function, and the target strength selected when generating the discrimination function are selected as the discrimination conditions. The sound identification condition setting support method according to claim 4.