JP7373358B2 - Sound extraction system and sound extraction method - Google Patents

Description

The present invention relates to a sound extraction system and a sound extraction method.

Conditions such as abnormalities in equipment or signs of failure are often manifested in sounds. Therefore, diagnosis based on the operating sounds of the equipment is important in order to understand the condition of the equipment. However, the diagnosis may be incorrect due to the influence of noise originating from sources other than the subject of diagnosis. Therefore, a sound extraction process is required that removes extraneous noise from the input signal and selectively extracts the sounds to be diagnosed.

Japanese Patent Laid-Open No. 2009-128906 (Patent Document 1) is a method for solving the problem of sound extraction. The publication describes a process of applying a constrained non-negative matrix factorization (NMF) to a mixed signal, the NMF being constrained by a denoising model, and the denoising model combining a training acoustic signal and a training noise signal. a training basis matrix and statistics of the weights of the training basis matrix, the application of which generates the weights of the basis matrix of the acoustic signal of the composite signal, and for reconstructing the acoustic signal; denoising a mixed signal comprising an acoustic signal and a noise signal, the method comprising: taking the product of a weight of a basis matrix of the acoustic signal and a training basis matrix of the training acoustic signal and the training noise signal; ``Method for.''

JP2009-128906A

The invention disclosed in Patent Document 1 uses NMF to separate a signal mixed with noise into speech and noise. However, it only works under conditions where training data for both the speech you want to extract and the noise you want to remove is given. For example, even if there is an abnormality in the operating sound of the equipment to be diagnosed, it is difficult to learn in advance the sound at the time of the abnormality, so the technology of Patent Document 1 cannot extract it. One way to reduce noise and make the sound of the diagnostic target louder is to place a microphone as close as possible to the diagnostic target and record the sound. However, even this is insufficient if the environmental noise is extremely large.

Therefore, an object of the present invention is to provide a sound extraction process that removes extraneous noise from an input signal and selectively extracts sounds to be diagnosed.

In order to solve the above problems, for example, the configurations described in the claims are adopted. In other words, multiple input sounds recorded at multiple locations with different distances from the diagnosis target are acquired in association with the distances, feature values are determined for each of the multiple input sounds, and multiple combinations of feature values and corresponding distances are obtained. This method is used to extract features of the sound to be diagnosed.

According to the present invention, extraneous noise can be removed from an input signal and sounds to be diagnosed can be selectively extracted. Problems, configurations, and effects other than those described above will be made clear by the following description of the embodiments.

FIG. 2 is an explanatory diagram of the hardware configuration of the embodiment. FIG. 3 is a functional block diagram related to processing during learning of a normal sound model according to the embodiment. FIG. 3 is a functional block diagram related to processing during execution of abnormality detection in the embodiment. 7 is a flowchart showing a processing procedure related to recording by a multi-distance input sound acquisition unit. FIG. 3 is a functional block diagram related to processing during learning of a normal sound model for each distance. FIG. 3 is a functional block diagram related to processing when performing abnormality detection by distance. 12 is a flowchart showing a sound extraction processing procedure by a sound extraction unit with distance-volume constraints. The flowchart which shows the 1st modification of sound extraction. The flowchart which shows the 2nd modification of sound extraction.

Examples will be described below with reference to the drawings.

FIG. 1 is an explanatory diagram of the hardware configuration of the embodiment. As shown in FIG. 1, the portable terminal 101 includes a microphone 102, an AD converter 103, and a distance measurement sensor 104, and is a terminal that a user can carry and move, such as a tablet terminal.

Microphone 102 sends an analog input signal to AD converter 103. The AD converter 103 converts the analog input signal into a digital output signal and sends it to the portable terminal 101. The distance sensor 104 measures the distance from the diagnosis target 105 and sends it to the portable terminal 101. If the distance measurement sensor 104 is not connected to the portable terminal 101, the user may measure the distance separately. Note that the diagnosis target 105 is, for example, a device installed as factory equipment.

The portable terminal 101 displays the indicated value of the distance from the diagnosis target 105 (short distance r1, long distance r2, etc.) and the current distance r from the diagnosis target 105 through its display. These displays allow the user to easily set the distance r to the indicated value. Also, if recording is stopped, the recording permission and recording start button are displayed. If the absolute value of the difference between the distance r and the instruction value is equal to or greater than the threshold value eps, a display indicating that recording is not possible is displayed. Otherwise, a display indicating that recording is possible is displayed, and the recording start button is enabled and can be pressed. This allows stable recording at accurate distances. After the start of recording and during recording, the total recording time T, the elapsed time t from the start of recording, and the remaining recording time (Tt) are displayed. This display has the effect of stabilizing the recording conditions and reducing the psychological burden on the user during recording.

The user performs recording by moving the portable terminal 101 to a plurality of positions having different distances from the diagnosis target 105, such as a short distance r1 and a long distance r2. This recording result includes the sound of the diagnosis target 105 and the background noise 106, and is used for learning a normal sound model of the sound of the diagnosis target 105 and detecting an abnormality of the diagnosis target 105.

Specifically, first, while the device that is the diagnosis target 105 is operating properly, recording is performed at short distance r1 and long distance r2, and the sound of the diagnosis target 105 is determined from the recording results at multiple distances. The features of the sound are extracted and learned as a normal sound model. After that, when performing abnormality detection of the diagnostic target 105, recording is performed again at short distance r1 and long distance r2, and the characteristics of the sound of diagnostic target 105 are extracted from the recording results at multiple distances, and the normal sound Anomalies are detected by comparing with the model.

FIG. 2 is a functional block diagram related to processing during learning of a normal sound model according to the embodiment. A series of processes during learning of a normal sound model may be performed on the portable terminal 101, or may be performed on another computer or server. The multiple distance input sound acquisition unit 201 acquires digital input sounds recorded at multiple distances from the diagnosis target 105 in association with the distances. A digital output signal output from the AD converter 103 is used as the digital input sound. The distance can be determined by using the output of the distance measuring sensor 104. For example, by moving the portable terminal 101 to multiple positions with different distances from the diagnosis target 105, such as short distance r1, long distance r2, etc., and having the user perform recording, multiple distances and digital input sound can be recorded. All you have to do is to associate it and get it.

The multiple distance input sound acquisition unit 201 outputs digital input sounds and distance time series. The digital input sound output by the multi-distance input sound acquisition unit 201 is a time domain signal that indicates signal values along the time axis. The distance time series is obtained by giving the distance at the time of recording as a value to the time axis of the digital input signal. Although the recording at the short distance r1 and the recording at the long distance r2 are performed separately and have no temporal continuity, the multiple distance input sound acquisition unit 201 continuously records the results of each recording to create one digital Output as input sound. At this time, the distance time series becomes one piece of data indicating each recording distance for a continuous time series.

The preprocessing unit 202 divides the digital input sound into frames, multiplies the frames by a window function, and performs short-time Fourier transform on the signal after the window function multiplication to calculate a frequency domain signal. The frequency domain signal is a set of M complex numbers, where if the frame size is N, one complex number corresponds to each of the (N/2+1)=M frequency bins. Furthermore, an input sound spectrogram (power spectrogram or amplitude spectrogram) is calculated from the frequency domain signal.

The distance-volume constrained sound extraction unit 203 extracts a diagnosis target extraction sound spectrogram based on the input sound spectrogram and the distance time series. The normal sound model learning unit 204 learns a model of the normal distribution of feature vectors consisting of consecutive L frames based on a large number of diagnostic target extracted sound spectrograms obtained in the past, and converts the model into a normal sound model. The information is stored in the database 205.

As a normal sound model, Gaussian mixture distribution (GMM), one-class support vector classifier, subspace method, local subspace method, k-means clustering, Deep Neural Network (DNN) autoencoder, Convolutional Neural Network (CNN) autoencoder, Long Short Term Memory (LSTM) autoencoder, variational autoencoder (VAE), etc. may be used.

An algorithm suitable for each normal sound model is known, and learning is performed using this algorithm. For example, in the case of GMM, the EM algorithm performs fitting using combinations of Gaussian distributions as many as the predetermined number of clusters. The learned normal sound model is defined by the calculated model parameters. All of the model parameters are stored in a normal sound model database (not shown).

FIG. 3 is a functional block diagram related to processing during execution of abnormality detection in the embodiment. A series of processes when performing abnormality detection are performed on the portable terminal 101. The processing from the multiple distance input sound acquisition unit 201 to the distance-volume constrained sound extraction unit 203 is the same as that in FIG.

The abnormality detection unit 301 reads a normal sound model from the normal sound model database 205 and performs abnormality detection processing on the extracted sound spectrogram to be diagnosed. That is, a time series of feature vectors consisting of consecutive L frames is calculated, and it is determined whether the time series can be generated from a normal sound model with sufficient probability.

ここで、

例えば、前記確率p(v | Θ)の負の対数尤度 - log p(v | Θ) を推定異常度と定義し、出力する。 For example, when the normal sound model is GMM, the M×L-dimensional feature vector v is the normal sound model (model parameters Θ=((μ1, Γ1, π1), ...(μq, Γq, πq), (μQ , ΓQ, πQ)) is calculated using the following equation.

here,

For example, the negative log likelihood of the probability p(v | Θ) - log p(v | Θ) is defined as the estimated abnormality degree and output.

When using a Deep Neural Network (DNN) autoencoder as a normal sound model, optimization algorithms such as SGD, Momentum SGD, AdaGrad, RMSprop, AdaDelta, Adam, etc. Internal parameters are optimized so that the reconstruction error between the feature vector and the output feature vector is reduced. When a feature vector of an abnormal sound is input, it is expected that the restoration error between them will become large. Therefore, the restoration error is defined as the estimated degree of abnormality and is output.

The abnormality display section 302 displays the value of the estimated abnormality degree, and further displays that there is an abnormality when the value of the estimated abnormality degree is equal to or higher than a certain value.

FIG. 4 shows an example of a flowchart showing a processing procedure related to recording by the multi-distance input sound acquisition unit 201. This example shows a case where the user is controlled to record at two distances, short distance r1 and long distance r2. You can also record it.

First, in S401, the multi-distance input sound acquisition unit 201 outputs an instruction to record at a short distance r1, and proceeds to S402.
In S402, the multiple distance input sound acquisition unit 201 assigns 0 to t, and proceeds to S403.
In S403, if t<T, the multi-distance input sound acquisition unit 201 proceeds to S404, otherwise proceeds to S408.
In S404, the distance measurement sensor 104 executes distance measurement to obtain the current distance r, and the process advances to S405.
In S405, the multiple distance input sound acquisition unit 201 proceeds to S406 if |rr1|<eps, otherwise returns to S401.
In S406, recording is performed using the microphone 102 and AD converter 103, and the process advances to S407.
In S407, the multiple distance input sound acquisition unit 201 adds the elapsed time Δt from the previous time to t, and returns to S403.

Next, in S408, the multiple distance input sound acquisition unit 201 instructs recording at long distance r2, and proceeds to S409.
In S409, the multiple distance input sound acquisition unit 201 assigns 0 to t, and proceeds to S410.
In S410, the multi-distance input sound acquisition unit 201 proceeds to S411 if t<T, otherwise ends the process.
In S411, the distance measurement sensor 104 executes distance measurement to obtain the current distance r, and the process advances to S412.
In S412, the multiple distance input sound acquisition unit 201 proceeds to S413 if |rr2|<eps, otherwise returns to S408.
In S413, recording is performed using the microphone 102 and AD converter 103, and the process advances to S414.
In S414, the multiple distance input sound acquisition unit 201 adds the elapsed time Δt from the previous time to t, and returns to S410.

Here, a modification example that can suppress the scale of the model parameters of the normal sound model will be disclosed. The difference from the configurations shown in FIGS. 2 and 3 is that a different normal sound model is learned for each distance, and the normal sound model corresponding to the distance at that time is used also in the abnormality detection process.

FIG. 5 is a functional block diagram related to processing during learning of a normal sound model for each distance. The processing from the multiple distance input sound acquisition unit 201 to the distance-volume constrained sound extraction unit 203 is the same as that in FIG. The distance-by-distance normal sound model learning unit 501 uses the diagnosis target extraction sound spectrogram output by the distance-volume constrained sound extraction unit 203 and the distance time series output by the multiple-distance input sound acquisition unit 201 to separately perform distance-by-distance normal sound model learning. normal sound model is learned and stored in the normal sound model database 502 for each distance. The learning algorithm for each model may be the same as that of the normal sound model learning section 204.

FIG. 6 is a block diagram related to processing when performing abnormality detection by distance. The processing from the multiple distance input sound acquisition unit 201 to the distance-volume constrained sound extraction unit 203 is the same as that in FIG. The distance-by-distance abnormality detection unit 601 uses the diagnosis target extraction sound spectrogram output by the distance-volume constrained sound extraction unit 203 and the distance time series output by the multiple distance input sound acquisition unit 201 to generate a diagnosis target extraction sound spectrogram. , the spectrogram is divided by distance at the time of recording, abnormality detection is performed using a normal sound model of the distance corresponding to the divided spectrogram, and the estimated degree of abnormality is output.

The integrated anomaly detection unit 602 integrates the estimated anomaly degrees between the divided spectrograms. If the AUC (Area under the curve) of the ROC (Receiver Operating Characteristic) curve for each distance d is calculated in advance as w_d, a weighting coefficient that increases as w_d increases is added to the abnormality degree of each distance d. The sum of the multiplied values is output as the integrated estimated abnormality degree. The weighting coefficient is, for example, −1.0/log(w_d).

The abnormality display section 603 displays the value of the integrated estimated abnormality degree, and further displays that there is an abnormality when the value of the integrated estimated abnormality degree is equal to or higher than a certain value.

FIG. 7 is a flowchart showing a sound extraction processing procedure by the distance-volume constrained sound extraction unit 203. First, in S701, the distance-volume constrained sound extraction unit 203 takes the input sound spectrogram as a matrix X, and initializes the unsupervised NMF for the matrix X. For example, the activation and base of each sound source are initialized with random numbers, and the process proceeds to S702.
In S702, the distance-volume constrained sound extraction unit 203 executes unsupervised NMF on the matrix X, and proceeds to S703.
In S703, the distance-volume constrained sound extraction unit 203 divides the activation obtained by the unsupervised NMF by distance d at the time of recording, and proceeds to S704. Activation indicates the time component in the input sound spectrogram, and the relationship between time and distance of the input sound is given as a distance time series. Therefore, it is possible to identify the time period corresponding to the distance from the distance time series and divide the activation into time periods corresponding to the distance.

In S704, the distance-volume constrained sound extraction unit 203 selects base k, and proceeds to S705.
In S705, the distance-volume constrained sound extraction unit 203 calculates the average activation value a_{k, d} over the divided time for the selected base k, and proceeds to S706. That is, a_{k, d} is the average value of the activation of base k in the time period recorded at distance d.
In S706, the distance-volume constrained sound extraction unit 203 determines whether the order of magnitude of a_{k, d} matches the order of magnitude of the reciprocal of distance 1/d. As a result of the determination, if they match (S706; Yes), the process advances to S707. If they do not match (S706; No), the process advances to S708.
In S707, the distance-volume constrained sound extraction unit 203 regards the selected base k as a component of the diagnosis target 105 and stores it in the set S, and proceeds to S708.
In S708, the distance-volume constrained sound extraction unit 203 determines whether all bases k have been selected. As a result of the determination, if unselected base k remains (S708; No), the process advances to S704. If all bases k have been selected (S708; Yes), the process advances to S709.

In S709, the distance-volume constrained sound extraction unit 203 restores the diagnosis target sound spectrogram. Specifically, the distance-volume constrained sound extraction unit 203 calculates the sum ^X of the multiplication W_k H_k of the activation W_k and the base vector H_k over all elements k of the set S. The distance-volume constrained sound extraction unit 203 outputs ^X as the diagnosis target extraction sound, and ends the process.

FIG. 8 is a flowchart showing a first modification of sound extraction. First, in S801, the distance-volume constrained sound extraction unit 203 initializes the unsupervised NMF for the matrix X_far, which is the spectrogram at the time of recording at the longest distance, and proceeds to S802.
In S802, the distance-volume constrained sound extraction unit 203 performs unsupervised NMF on the matrix X_far, outputs the initial solution W_far_ini of background noise activation and the background noise basis vector H_far_ini for the matrix X_far, and proceeds to S803.

In S803, the distance-volume constrained sound extraction unit 203 initializes a semi-supervised NMF for the matrix X that is the input sound spectrogram. That is, as background noise activation, the initial solution at the time when the recording was made at the longest distance is set as W_far_ini, and the initial solution at other times is set as the time average value of W_far_ini. Store H_far_ini in the base vector of background noise. As the activation of the sound to be diagnosed, a sufficiently small positive value is stored in the initial solution at the time when the recording was made at the longest distance, and random numbers are stored in the initial solutions at other times. A random number is stored as an initial solution in the basis vector of the sound to be diagnosed, and the process advances to S804.

In S804, the distance-volume constrained sound extraction unit 203 executes semi-supervised NMF on the matrix X, and proceeds to S805.
In S805, the distance-volume constrained sound extraction unit 203 restores the diagnosis target sound spectrogram in the same manner as in S709 of FIG. 7, and ends the process.

FIG. 9 is a flowchart showing a second modification of sound extraction. 9 is the same as FIG. 8 except that semi-supervised NMF with distance regularization is performed on the matrix X in S901 corresponding to S804 in FIG. Distance regularization is a process in which activation is multiplied by a constant for each distance so that the ratio of the average value of activation at times when the distance is r times becomes 1/r for each NMF repetition process.

As described above, the sound extraction system and sound extraction method according to the present embodiment acquires a plurality of input sounds recorded at a plurality of positions having different distances from the diagnosis target and associates them with distance, and extracts the plurality of input sounds. A feature quantity is obtained for each, and a plurality of combinations of the feature quantity and the corresponding distance are used to extract the feature quantity of the sound to be diagnosed. Therefore, extraneous noise can be removed from the input signal and sounds to be diagnosed can be selectively extracted.

The sound extraction system and sound extraction method according to the present embodiment do not require the condition that learning data for both the sound to be extracted and the noise to be removed are given in advance. For example, the sounds that can be learned in advance may be only sounds that are subject to diagnosis during normal conditions and are mixed with noise. In addition, it is possible to extract not only the normal sounds of the diagnostic target but also abnormal sounds that are difficult to learn in advance.

Further, the sound extraction system and sound extraction method according to the present embodiment can extract a sound to be diagnosed using a plurality of input sounds recorded at a plurality of positions by moving the same microphone. Therefore, it is possible to operate the portable terminal device alone. Further, by outputting the relationship between the position where recording should be performed and the position of the microphone and guiding the user, it is possible to stabilize the recording conditions and reduce the psychological burden on the user during recording.

Further, the sound extraction system and sound extraction method according to the present embodiment can detect an abnormality in the diagnosis target by calculating the estimated degree of abnormality based on the extracted feature amount of the sound of the diagnosis target. Further, the display according to the estimated degree of abnormality allows the user to recognize the abnormality. The degree of abnormality estimation may be determined, for example, by comparing the extracted feature amount of the sound of the diagnosis target with the feature amount of the sound when the diagnosis target is operating properly. Furthermore, it is also possible to learn in advance the feature amount of sound during proper operation of the diagnostic target as a normal sound model.

In addition, the sound extraction system and sound extraction method according to the present embodiment can extract feature components that are common to a plurality of feature quantities corresponding to a plurality of distances and exhibit a magnitude relationship according to the distances as a feature of the sound to be diagnosed. It can be extracted as a quantity. In other words, since recording is performed at multiple distances, it is possible to identify components that are high in volume at times when they are recorded at short distances and low in volume at times when they are recorded at long distances as diagnostic target sounds, and extract only those components. can do. Furthermore, it is also possible to extract characteristic components that show not only a magnitude relationship but also a change according to distance as a characteristic amount of the sound to be diagnosed.

Furthermore, the sound extraction system and sound extraction method according to the present embodiment can calculate frequency domain signals for each of a plurality of input sounds, calculate a spectrogram from each frequency domain signal, and use the spectrogram as a feature quantity. can.

Although the description is omitted in this embodiment, the diagnosis target may be a device having multiple operation modes. In this case, recording is performed while the device is operating properly in one of a plurality of operating modes, and a normal sound model is learned.

Further, although the present embodiment has been described including learning of a normal sound model, the normal sound model may be given in advance. In addition, in this embodiment, the configuration was explained by exemplifying the configuration in which the diagnosis target sound is extracted and then the abnormality is detected. However, for example, the diagnosis target sound is reconstructed from the extracted diagnosis target sound spectrogram, It is also possible for the user to listen to the configured sounds to be diagnosed and determine the presence or absence of an abnormality. Further, in this embodiment, a case is illustrated in which recording is performed at each specified distance, but the sound to be diagnosed may be extracted by moving the recording position while continuing recording.

Note that the present invention is not limited to the above-described embodiments, and includes various modifications. For example, the embodiments described above are described in detail to explain the present invention in an easy-to-understand manner, and the present invention is not necessarily limited to having all the configurations described. Furthermore, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. Further, it is possible to add, delete, or replace a part of the configuration of each embodiment with other configurations.

Further, each of the above-mentioned configurations, functions, processing units, processing means, etc. may be partially or entirely realized in hardware by designing, for example, an integrated circuit. Furthermore, each of the above configurations, functions, etc. may be realized by software by a processor interpreting and executing a program for realizing each function. Information such as programs, tables, files, etc. that realize each function can be stored in a memory, a recording device such as a hard disk, an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, or a DVD.

Further, the control lines and information lines are shown to be necessary for explanation purposes, and not all control lines and information lines are necessarily shown in the product. In reality, almost all components may be considered to be interconnected.

101... Portable terminal, 102... Microphone, 103... AD converter, 104... Distance sensor, 105... Diagnosis target, 106... Background noise, 201... Multiple distances Input sound acquisition unit, 202... Preprocessing unit, 203... Sound extraction unit with distance-volume constraints, 204... Normal sound model learning unit, 205... Normal sound model database, 301... Abnormality Detection unit, 302... Abnormality display unit, 501... Distance normal sound model learning unit, 502... Distance normal sound model database, 601... Distance based abnormality detection unit, 602... Integrated abnormality Detection unit, 603... Abnormality display unit

Claims (10)

a multi-distance input sound acquisition unit that acquires a plurality of input sounds recorded at a plurality of positions having different distances from the diagnosis target in association with the distances;
a preprocessing unit that calculates feature amounts for each of the plurality of input sounds;
a sound extraction unit that extracts a feature amount of the sound to be diagnosed using a plurality of combinations of the feature amount and corresponding distance;
Equipped with
The sound extraction system is characterized in that the multiple distance input sound acquisition unit acquires a plurality of input sounds recorded at the plurality of positions by moving the same microphone. The multi-distance input sound acquisition unit outputs the relationship between the microphone and the plurality of positions, guides the microphone to the plurality of positions, and moves the microphone to collect the plurality of input sounds recorded at the plurality of positions. The sound extraction system according to claim 1, characterized in that the sound extraction system acquires. 2. The apparatus according to claim 1, further comprising an abnormality detection section that calculates an estimated degree of abnormality indicating an abnormality of the diagnosis target based on the feature amount of the sound of the diagnosis target extracted by the sound extraction section. sound extraction system. an anomaly detection unit that calculates an estimated abnormality degree indicating an abnormality of the diagnosis target based on the feature amount of the sound of the diagnosis target extracted by the sound extraction unit;
The sound extraction system according to claim 1, further comprising: an abnormality display unit that displays a display according to the estimated abnormality degree. anomaly detection that calculates an estimated degree of abnormality indicating an abnormality in the diagnosis target by comparing the feature amount of the sound of the diagnosis target extracted by the sound extraction unit with the feature amount of the sound when the diagnosis target is operating properly; The sound extraction system according to claim 1, further comprising: a. a learning unit that learns the feature amount of sound when the diagnosis target is operating properly;
an anomaly detection unit that calculates an estimated degree of abnormality indicating an abnormality of the diagnosis target by comparing the feature amount of the sound of the diagnosis target extracted by the sound extraction unit and the feature amount of the sound during the proper operation; The sound extraction system according to claim 1, further comprising: a sound extraction system according to claim 1; The sound extraction unit extracts a feature component that exists in common among the plurality of feature quantities corresponding to the plurality of distances and exhibits a magnitude relationship according to the distance, as a feature quantity of the sound to be diagnosed. The sound extraction system according to claim 1. The sound extraction unit is characterized in that it extracts a feature component that exists in common among the plurality of feature quantities corresponding to the plurality of distances and shows a change according to the distance, as a feature quantity of the sound to be diagnosed. The sound extraction system according to claim 1. The preprocessing unit calculates frequency domain signals for each of the plurality of input sounds, calculates a spectrogram from each frequency domain signal, and uses the spectrogram as the feature amount. sound extraction system. a multi-distance input sound acquisition step of acquiring a plurality of input sounds recorded at a plurality of positions having different distances from the diagnosis target in association with the distances;
a preprocessing step of calculating feature amounts for each of the plurality of input sounds;
a sound extraction step of extracting a feature amount of the sound to be diagnosed using a plurality of combinations of the feature amount and corresponding distance ;
The sound extraction method is characterized in that the plural distance input sound acquisition step moves the same microphone to acquire a plurality of input sounds recorded at the plurality of positions.

Images