WO2015060375A1 - Biological sound signal processing device, biological sound signal processing method, and biological sound signal processing program - Google Patents

Biological sound signal processing device, biological sound signal processing method, and biological sound signal processing program Download PDF

Info

Publication number
WO2015060375A1
WO2015060375A1 PCT/JP2014/078182 JP2014078182W WO2015060375A1 WO 2015060375 A1 WO2015060375 A1 WO 2015060375A1 JP 2014078182 W JP2014078182 W JP 2014078182W WO 2015060375 A1 WO2015060375 A1 WO 2015060375A1
Authority
WO
WIPO (PCT)
Prior art keywords
sound
matrix
biological
sound signal
continuous
Prior art date
Application number
PCT/JP2014/078182
Other languages
French (fr)
Japanese (ja)
Inventor
智弥 酒井
千弥 善安
末治 宮原
Original Assignee
国立大学法人 長崎大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国立大学法人 長崎大学 filed Critical 国立大学法人 長崎大学
Priority to JP2015543900A priority Critical patent/JPWO2015060375A1/en
Publication of WO2015060375A1 publication Critical patent/WO2015060375A1/en

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B7/00—Instruments for auscultation
    • A61B7/02—Stethoscopes
    • A61B7/04—Electric stethoscopes
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B7/00—Instruments for auscultation
    • A61B7/003—Detecting lung or respiration noise

Definitions

  • the present invention relates to a biological sound signal processing device, a biological sound signal processing method, and a biological sound signal processing program for processing biological sounds such as lung sounds.
  • a diagnosis support device has been developed to convert lung sound signals (pulmonary sound signals) obtained by electronic stethoscopes into digital data, analyze the data, and use the analysis results for diagnosis (sound diagnosis). Is underway.
  • Lung sounds are roughly divided into respiratory sounds and auxiliary noises as abnormal sounds.
  • the sub-noise is further divided into a ra sound and others, and the ra sound is further divided into an intermittent ra sound and a continuous ra sound.
  • Intermittent rales include water bubbles and haircut sounds, and continuous rales include whistle sounds and snoring sounds.
  • a method in which lung sounds are classified into normal breath sounds and continuous rales using fast Fourier transform and its inverse transform (see, for example, Patent Document 1).
  • this method first, an amplitude spectrum and a power spectrum are calculated by performing a fast Fourier transform (FFT) on a time waveform of a lung sound.
  • FFT fast Fourier transform
  • an inverse FFT process is performed on the amplitude spectrum at a point where the local dispersion value of the power spectrum exceeds the threshold value. In this way, normal breath sounds and continuous rales can be distinguished.
  • Non-Patent Document 1 a technique for separating respiratory sounds and intermittent rales from lung sounds is known (for example, see Non-Patent Document 1).
  • lung sounds are separated based on a sparse expression that constitutes the lung sounds most simply by the sum of a Fourier transform signal and a wavelet signal.
  • This sparse representation f (t) is classified as a breathing sound, and w (t) is classified as an intermittent rale.
  • the method of separating lung sounds into normal breath sounds and continuous rales using fast Fourier transform and its inverse transform is a kind of filtering in which frequency bands are adaptively selected. For this reason, when both share the same frequency component, it cannot be distinguished.
  • the signal that is used as the normal breathing sound here may include an intermittent sound that is an abnormal sound.
  • the original lung sound signal contains continuous rales. If there is, the continuous rarity is not always classified into either f (t) or w (t). For this reason, it cannot respond to the analysis process of continuous rales.
  • lung sound signals including various abnormal sounds cannot be accurately classified.
  • an object of the present invention is to enable more accurate separation of continuous and intermittent rales from lung sounds of humans and the like.
  • the present invention provides a robust principal component analysis unit that decomposes an original sound matrix representing an original sound spectrogram of a biological sound signal into a sparse matrix and a low rank matrix by robust principal component analysis in the biological sound signal processing apparatus.
  • a continuous sound processing unit that converts the sparse matrix to obtain continuous biological sound from the biological sound signal, and a biological body that converts the low rank matrix and excludes continuous biological sound from the biological sound signal
  • a discontinuous sound processing unit that obtains sound.
  • the present invention also relates to a biological sound signal processing method, a robust principal component analysis step of decomposing an original sound matrix representing an original sound spectrogram of a biological sound into a sparse matrix and a low rank matrix by robust principal component analysis, and converting the sparse matrix
  • a robust principal component analysis step of decomposing an original sound matrix representing an original sound spectrogram of a biological sound into a sparse matrix and a low rank matrix by robust principal component analysis, and converting the sparse matrix
  • obtaining a third step is
  • the present invention provides a robust principal component analysis means for decomposing an original sound matrix representing an original sound spectrogram of biological sound into a sparse matrix and a low rank matrix by robust principal component analysis, and the sparse A continuous sound processing means for converting a matrix to obtain continuous biological sounds from the biological sound signal; and a non-continuous sound processing means for obtaining biological sounds excluding continuous biological sounds from the biological sound signals by converting the low rank matrix It is characterized by functioning as a continuous sound processing means.
  • FIG. 1 is a block diagram of a first embodiment of a biological sound signal processing device according to the present invention. It is a flowchart of the biological sound signal processing method in 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention. It is a graph of the original sound which processes in the 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention. It is the figure which represented the value of the element of the matrix which has the value of the original sound spectrogram obtained by carrying out the short-time Fourier transform of the original sound signal in the 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention by the light and shade. .
  • FIG. 1 is a block diagram of a first embodiment of a biological sound signal processing apparatus according to the present invention.
  • the biological sound signal processing device 90 includes a preliminary processing unit 10, a robust principal component analysis unit 40, a sparse matrix storage unit 22, a low rank matrix storage unit 32, a continuous sound processing unit 20, and a discontinuous sound processing unit 30. is doing.
  • the biological sound signal processing device 90 is constructed on, for example, one computer.
  • the biological sound signal processing device 90 may be constructed on a plurality of computers connected by a network.
  • the preliminary processing unit 10 includes an input unit 12, a Fourier transform unit 14, and a matrix generation unit 16.
  • the continuous sound processing unit 20 includes a continuous sound spectrogram generation unit 24 and a first inverse Fourier transform unit 26.
  • the discontinuous sound processing unit 30 includes a discontinuous sound spectrogram generation unit 34, a second inverse Fourier transform unit 36, and a signal extraction unit 38.
  • the preliminary processing unit 10 generates an original sound matrix that represents an original sound spectrogram of a biological sound.
  • the biological sound is detected by a biological sound detection device such as an electronic stethoscope (not shown), and is supplied to the biological sound signal processing device 90 as an electrical signal.
  • the robust principal component analysis unit 40 performs a robust principal component analysis on the original sound matrix generated by the preliminary processing unit 10 to obtain a sparse matrix and a low rank matrix.
  • the sparse matrix is stored in the sparse matrix storage unit 22.
  • the low rank matrix is stored in the low rank matrix storage unit 32.
  • the continuity sound processing unit 20 processes the sparse matrix stored in the sparse matrix storage unit 22 to generate a continuity sound in the original sound.
  • the discontinuous sound processing unit 30 processes the low rank matrix stored in the low rank matrix storage unit 32 to generate a discontinuous sound in the original sound.
  • the continuous sound and the discontinuous sound generated by the continuous sound processing unit 20 and the discontinuous sound processing unit 30 are, for example, subjected to D / A conversion and output by a speaker (not shown).
  • the waveform of continuous sound and discontinuous sound may be displayed on the display.
  • a continuity sound signal and a discontinuity sound signal may be transmitted to an external device, and abnormality detection or the like may be performed by the external device.
  • FIG. 2 is a flowchart of the biological sound signal processing method in the present embodiment.
  • the input unit 12 captures a signal that detects a body sound from a body sound detection device such as an electronic stethoscope (not shown).
  • the detected biological sound is called an original sound.
  • a signal that electrically represents the original sound is referred to as an original sound signal s (t).
  • the biological sound is, for example, a human lung sound.
  • the input unit 12 performs A / D conversion to convert the original sound signal into digital data.
  • FIG. 3 is a graph of lung sound signals to be processed in the present embodiment.
  • the horizontal axis represents elapsed time (seconds), and the vertical axis represents signal intensity (amplitude).
  • the lung sound recorded on the 60th track of the CD in the appendix of Non-Patent Document 2 is used as the original sound.
  • the Fourier transform unit 14 performs a short-time Fourier transform on the original sound signal s (t) to obtain a complex sound spectrogram (hereinafter referred to as a spectrogram) represented by a complex quantity in the time-frequency domain (step 1).
  • the original sound spectrogram S ( ⁇ , t) is obtained by subjecting the original sound signal s (t) multiplied by the time window function to a discrete Fourier transform.
  • the original sound spectrogram S ( ⁇ , t) has a complex value and represents the amplitude and phase of the component of the angular frequency ⁇ constituting the signal at time t representing the position of the time window function.
  • Time t takes a discrete value with a time width ⁇ t2 for shifting the time window function. It is assumed that the time width ⁇ t2 for shifting the time window function does not exceed the time window width ⁇ t1. That is, ⁇ t2 ⁇ t1. Further, the angular frequency ⁇ is discretized at an interval proportional to the reciprocal of the time window width ⁇ t1.
  • the matrix generation unit 16 creates an original sound matrix D having the amplitude
  • the row number i and the column number j of the original sound matrix D correspond to the i-th angular frequency ⁇ i and the j-th time t j .
  • An element Dij in the i-th row and j-th column of the matrix D is an absolute value of a complex number S ( ⁇ i , t j ) constituting the original sound spectrogram S ( ⁇ , t).
  • FIG. 4 is a diagram in which the values of the elements of the original sound matrix having the values of the original sound spectrogram obtained by performing a short-time Fourier transform on the original sound signal in this embodiment are represented by shading.
  • the robust principal component analysis unit 40 decomposes the original sound matrix D so as to be the sum of the low rank matrix A and the sparse matrix E.
  • such matrix decomposition is called robust principal component analysis.
  • Non-Patent Document 4 proposes a fast convergence algorithm improved from the extended Lagrangian method.
  • a given matrix is approximated by a low rank matrix.
  • the low rank matrix is constructed by the product of eigenvectors associated with only the main eigenvalues (or singular values) of a given matrix.
  • a given original sound matrix D is approximated by a low rank matrix A while allowing only a part of its elements to be modified.
  • the original sound matrix D is decomposed into a sum of a low rank matrix A and a sparse matrix E representing a correction amount.
  • the rank (rank) of the matrix A and the number of elements to be corrected are both as small as possible.
  • the uniqueness of the solution in this case is disclosed in Non-Patent Document 3.
  • FIG. 5 is a diagram in which the values of the elements of the low rank matrix obtained by the robust principal component analysis in the present embodiment are represented by shading.
  • FIG. 6 is a diagram in which the values of the elements of the sparse matrix obtained by the robust principal component analysis in the present embodiment are represented by shading.
  • the low rank matrix A has elements of a matrix D that can be easily constructed by the product of main eigenvectors. Therefore, the pattern in FIG. 5 exhibited by the rows or columns of the low rank matrix A tends to appear in a plurality of similar patterns, such as the vertical stripes and horizontal stripe patterns shown in FIG. 4 in which the values of the elements of the original sound matrix are represented by shading. is there.
  • the sparse matrix E has the components excluded to approximate the original sound matrix D by such a low rank matrix A as elements. Therefore, as shown in FIG. 6, the sparse matrix E exhibits an arbitrary curved or spotted pattern having no regularity such as vertical stripes and horizontal stripes.
  • step 4 a continuous sound spectrogram E ( ⁇ , t) corresponding to the sparse matrix E obtained in step 3 is generated.
  • the continuity sound spectrogram E ( ⁇ , t) is generated by the continuity sound spectrogram generation unit 24 reading the sparse matrix E from the sparse matrix storage unit 22.
  • the complex number E ( ⁇ i , t j ) constituting the continuity sound spectrogram E ( ⁇ , t) is a complex number having an element of the sparse matrix as an amplitude and an argument of the original sound spectrogram as an argument.
  • the complex number E ( ⁇ i , t j ) is the argument ⁇ ij of the complex number S ( ⁇ i , t j ) that constitutes the element E ij in the i-th row and j-th column of the matrix E and the original sound spectrogram S ( ⁇ , t). Is obtained by the following equation.
  • the first inverse Fourier transform unit 26 performs a short-time inverse Fourier transform on the continuous sound spectrogram E ( ⁇ , t) to obtain a lung sound signal e (t). More specifically, continuity sound spectrogram E (omega, t) complex spectrum E (omega, t j) for each time t j of the inverse Fourier transform short time, in the time window function at each time t j A lung sound signal is obtained.
  • the lung sound signal e (t) is obtained by averaging the lung sound signals according to the overlap of the time window functions.
  • FIG. 7 is an e (t) graph of a lung sound signal obtained by performing an inverse Fourier transform on a sparse matrix in the present embodiment.
  • FIG. 7 shows the lung sound signal e (t) output from the first inverse Fourier transform unit 26 as digital data, which is simulated and displayed like an analog signal.
  • the lung sound signal e (t) is D / A converted by a biological sound output unit (not shown) and reproduced as a lung sound by a speaker (not shown) or the like, it becomes a continuous sound.
  • the original sound signal s (t) used as an input of processing is a lung sound including a continuous ra sound (high-pitched whistle sound) adopted from Non-Patent Document 2 from the 60th track of the CD of the appendix. It is.
  • the continuous ra sound has a curved pattern in FIG. 4 representing the spectrogram of the original sound signal s (t), and is clearly separated into the spectrogram represented by the sparse matrix E shown in FIG. From this, it can be seen that the continuous sound is well separated as the lung sound signal e (t) in step 5 by the processing from step 1 to step 5 described above.
  • the continuous rales used in this example are high-pitched whistle sounds, but low-pitched snoring sounds can also be separated. Similar to the present embodiment, if the spectrogram of sounds other than continuous rales (breathing sounds and intermittent rachunes) is easily represented by the low rank matrix A, the low rank matrix A in the robust principal component analysis of step 3 is performed. The spectrogram of the continuous ra-tone is separated as a component (sparse matrix E) excluded to approximate the matrix D.
  • step 6 the discontinuous sound spectrogram generation unit 34 generates the discontinuous sound spectrogram A ( ⁇ , t) corresponding to the low rank matrix A separated in step 3.
  • the generation method is the same as in Step 4. That is, the complex number A ( ⁇ i , t j ) constituting the discontinuous sound spectrogram A ( ⁇ , t) is a complex number having an element of the low rank matrix as an amplitude and an argument of the original sound spectrogram as an argument.
  • a ( ⁇ i , t j ) A ij (cos ( ⁇ ij ) + isin ( ⁇ ij ))
  • step 7 the second inverse Fourier transform unit 36 performs a short-time inverse Fourier transform on the discontinuous sound spectrogram A ( ⁇ , t) to obtain a lung sound signal a (t).
  • This method is the same as step 5. That is, discontinuous sound spectrogram A (omega, t) complex spectrum A (omega, t j) for each time t j of the inverse Fourier transform short time, lung sounds signals within a time window function at each time t j Get.
  • a lung sound signal a (t) is obtained by averaging the lung sound signals according to the overlap of the window functions.
  • FIG. 8 is a graph of the lung sound signal a (t) obtained by performing inverse Fourier transform on the low rank matrix A in the present embodiment.
  • FIG. 8 shows the lung sound signal output from the second inverse Fourier transform unit 36 as digital data by simulating and displaying it as an analog signal.
  • the lung sound signal a (t) is D / A converted by a biological sound output unit (not shown) and reproduced as a lung sound by a speaker or the like (not shown), a sound obtained by removing the continuous sound from the original sound signal s (t) is obtained.
  • the signal extraction unit 38 further extracts the lung sound signal f (t) and the lung sound signal w (t) from the lung sound signal e (t) by the method of Non-Patent Document 1.
  • the method of Non-Patent Document 1 extracts lung sounds based on sparse representation.
  • the lung sound signal a (t) is expressed by the sum of the Fourier synthesized signal (f (t)) and the wavelet synthesized signal (w (t)).
  • the number of non-zero Fourier components and the number of non-zero wavelet components are both as small as possible.
  • Non-Patent Document 1 discloses an algorithm for this purpose.
  • FIG. 9 is a graph of a lung sound signal f (t) that is a Fourier component extracted from the lung sound signal e (t) in the present embodiment.
  • FIG. 9 shows the lung sound signal f (t) output from the signal extraction unit 38 as digital data by simulating it as an analog signal.
  • the lung sound signal f (t) is D / A converted by a biological sound output unit (not shown) and reproduced as a lung sound by a speaker (not shown), a respiratory sound in the original sound signal s (t) is output.
  • FIG. 10 is a graph of the lung sound signal w (t), which is a wavelet component extracted from the lung sound signal e (t) in the present embodiment.
  • FIG. 10 shows the lung sound signal w (t) output from the signal extraction unit 38 as digital data by simulating it as an analog signal.
  • the lung sound signal w (t) is D / A converted by a biological sound output unit (not shown) and reproduced as a lung sound by a speaker (not shown) or the like, an intermittent ra sound in the original sound signal s (t) is output.
  • the original sound signal s (t) is continuous sound data (high-pitched whistle sound) adopted from the 60th track in the CD of Appendix of Non-Patent Document 2.
  • the waveform of the original sound signal s (t) shown in FIG. 3 that intermittent sound is mixed in addition to continuous rarity.
  • the spectrogram representing the original sound signal s (t) shown in FIG. 4 in addition to the continuous ra sound that exhibits a curvilinear pattern, an intermittent sound that exhibits a vertical stripe pattern (intermittent ra sound), a low frequency The presence of sound (breathing sound) that continues in the belt can be confirmed.
  • respiratory sounds, continuous rales, and intermittent rales are separated and extracted from human lung sounds.
  • sounds similar to a natural ra sound and an intermittent ra sound are included, those sounds can be separated and extracted as in the present embodiment.
  • FIG. 11 is a block diagram of a second embodiment of the biological sound signal processing device according to the present invention.
  • FIG. 12 is a flowchart of the biological sound signal processing method in the present embodiment.
  • This embodiment is different from the first embodiment in that a complex matrix is used as a whole process. For this reason, the matrix generation part 16 in 1st Embodiment does not exist in the biological sound signal processing apparatus of this Embodiment.
  • step 2 in the first embodiment is omitted, and in step 3, the robust principal component analysis unit 40 directly performs a robust principal component analysis on the original sound spectrogram S ( ⁇ , t) to obtain a low rank matrix A and Get the sparse matrix E.
  • Steps 5 and 7 for obtaining a separated lung sound signal, and step 8 and subsequent steps are the same as those in the first embodiment.
  • the step of converting the complex matrix in the first embodiment into a real matrix and converting it again into a complex matrix after robust principal component analysis is omitted in this embodiment. Since the matrix of complex numbers representing the amplitude and the phase is separated, the separation performance can be improved as compared with the first embodiment in which only the amplitude is separated. Note that, since the matrix to be subjected to the robust principal component analysis is a complex matrix, the calculation time is slightly longer than that of the first embodiment. However, if an appropriate solution is used, it is sufficiently practical.
  • FIG. 13 is a block diagram of a third embodiment of the biological sound signal processing device according to the present invention.
  • FIG. 14 is a flowchart of the biological sound signal processing method in the present embodiment.
  • This embodiment is different from the first embodiment in that short-time cosine transform and short-time inverse cosine transform are used instead of short-time Fourier transform and short-time inverse Fourier transform.
  • a short-time cosine transform unit 41 is provided instead of the short-time Fourier transform unit 14 in the first embodiment.
  • This embodiment is different from the first embodiment in that a real number matrix is used as a whole process. For this reason, the matrix generation part 16 in 1st Embodiment does not exist in the biological sound signal processing apparatus of this Embodiment.
  • step 2 in the first embodiment is omitted, and in step 3, the robust principal component analysis unit 50 directly robusts the real original spectrogram S ( ⁇ , t) obtained by the short-time cosine transform unit. Principal component analysis is performed to obtain a low rank matrix A and a sparse matrix E.
  • a short-time inverse cosine transform unit 52 is provided instead of the first short-time inverse Fourier transform unit 26 in the first embodiment.
  • a short-time inverse cosine transform unit 53 is provided instead of the second short-time inverse Fourier transform unit 36 in the first embodiment.
  • the cosine transform has the advantage that it can be processed in half the storage area compared to the Fourier transform of the real signal. However, since the process assumes that the signal is an even function, the separation performance may be slightly inferior to that of the other embodiments.
  • A UKV T is obtained from the low rank matrix A by singular value decomposition.
  • a matrix U, a matrix K, and a matrix V are obtained. If the size of the low-rank matrix A is m ⁇ n and the rank (rank) is r, the matrix K is a diagonal r-order square matrix with singular values on the diagonal, and the matrices U and V are r An m ⁇ r matrix and an n ⁇ r matrix composed of a left singular vector and a right singular vector Note that V T is a transposed matrix of V when the matrix V is a real matrix, and a conjugate transposed matrix of V when the matrix V is a complex matrix.
  • the left singular vector is a basis on which the column vector of the matrix A can be synthesized.
  • the column vector of the low rank matrix A represents the instantaneous frequency spectrum of sounds other than continuous biological sounds, particularly breathing sounds and intermittent sounds.
  • the left singular vector is the basis for constructing the instantaneous frequency spectrum of these sounds.
  • the right singular vector indicates the breakdown of the instantaneous frequency spectrum at an arbitrary time. That is, the component of the j-th right singular vector represents how much the base of the j-th left singular vector appears at each time.
  • singular value decomposition is a method of matrix decomposition for obtaining principal components for a set of row vectors and a set of column vectors constituting a matrix.
  • a singular value representing the size of the principal component and an orthonormal basis representing the direction of the principal component are obtained.
  • the singular vector associated with a singular value of zero is not uniquely determined.
  • Corresponding singular vectors on the left and right are arbitrary in terms of a code or a complex multiple of size 1.
  • the low rank matrix A obtained in the present embodiment is uniquely determined because it is composed of non-zero singular values and the left and right singular vectors associated therewith.
  • SYMBOLS 10 Preliminary processing part, 12 ... Input part, 14 ... Fourier transform part, 16 ... Matrix generation part, 20 ... Continuous sound processing part, 22 ... Sparse matrix storage part, 24 ... Continuous sound spectrogram generation part, 26 ... 1st DESCRIPTION OF SYMBOLS 1 Inverse Fourier-transform part, 30 ... Discontinuous sound processing part, 32 ... Low rank matrix storage part, 34 ... Discontinuous sound spectrogram generation part, 36 ... 2nd inverse Fourier transform part, 38 ... Signal extraction part, 40 ... Robust principal component analysis unit, 41 ... cosine transform unit, 50 ... robust principal component analysis unit, 52 ... inverse cosine transform unit, 53 ... inverse cosine transform unit, 90 ... biological sound signal processing device

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Pulmonology (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)

Abstract

[Problem] To make it possible to more accurately distinguish continuous rhonchi from discontinuous rales in lung sounds of humans, etc. [Solution] The biological sound detection signal processing device (90) comprises a robust principal component analysis unit (40), a continuous sound processing unit (20), and a discontinuous sound processing unit (30). The robust principal component analysis unit (40) receives original biological sounds with an input section (12), performs Fourier transforms in a Fourier transform section (14), and performs robust principal component analysis of original sound matrices generated by a matrix-generating section (16). When a sparse matrix obtained from robust principal component analysis is processed with the continuous sound processing unit (20), continuous biological sounds are obtained from the original sounds. When a low rank matrix obtained from robust principal component analysis is processed with the discontinuous sound processing unit (30), biological sounds in which continuous biological sounds have been excluded from the original sounds are obtained.

Description

生体音信号処理装置、生体音信号処理方法および生体音信号処理プログラムBiological sound signal processing apparatus, biological sound signal processing method, and biological sound signal processing program
　本発明は、肺音などの生体音を処理する生体音信号処理装置、生体音信号処理方法および生体音信号処理プログラムに関する。 The present invention relates to a biological sound signal processing device, a biological sound signal processing method, and a biological sound signal processing program for processing biological sounds such as lung sounds.
　近年、電子聴診器などにより取得される肺音の信号（肺音信号）などをデジタルデータに変換してデータ解析し、その解析結果を診断（音診断）に活用するための診断支援装置の開発が進められている。 In recent years, a diagnosis support device has been developed to convert lung sound signals (pulmonary sound signals) obtained by electronic stethoscopes into digital data, analyze the data, and use the analysis results for diagnosis (sound diagnosis). Is underway.
　肺音は、呼吸音と異常音としての副雑音に大別される。副雑音はさらに、ラ音とその他に分けられ、ラ音はさらに断続性ラ音と連続性ラ音に分けられる。断続性ラ音には水泡音と捻髪音が含まれ、連続性ラ音には笛音といびき音が含まれる。 Lung sounds are roughly divided into respiratory sounds and auxiliary noises as abnormal sounds. The sub-noise is further divided into a ra sound and others, and the ra sound is further divided into an intermittent ra sound and a continuous ra sound. Intermittent rales include water bubbles and haircut sounds, and continuous rales include whistle sounds and snoring sounds.
　肺音を高速フーリエ変換およびその逆変換を用いて正常呼吸音と連続性ラ音に分別する方法が知られている（たとえば特許文献１参照）。この方法では、まず、肺音の時間波形を高速フーリエ変換（ＦＦＴ）して振幅スペクトルおよびパワースペクトルを算出する。次に、このパワースペクトルの局所分散値が閾値を越えた点における振幅スペクトルを逆ＦＦＴ処理する。このようにして、正常呼吸音と連続性ラ音が分別できる。 A method is known in which lung sounds are classified into normal breath sounds and continuous rales using fast Fourier transform and its inverse transform (see, for example, Patent Document 1). In this method, first, an amplitude spectrum and a power spectrum are calculated by performing a fast Fourier transform (FFT) on a time waveform of a lung sound. Next, an inverse FFT process is performed on the amplitude spectrum at a point where the local dispersion value of the power spectrum exceeds the threshold value. In this way, normal breath sounds and continuous rales can be distinguished.
　また、肺音から呼吸音と断続性ラ音を分別する技術が知られている（たとえば非特許文献１参照）。この技術では、肺音信号をフーリエ変換信号とウェーブレット信号との和で最も簡潔に肺音を構成するスパース（ｓｐａｒｓｅ）表現に基づき肺音を分離するものである。このスパース表現ｆ（ｔ）が呼吸音、ｗ（ｔ）が断続性ラ音として分別される。 Also, a technique for separating respiratory sounds and intermittent rales from lung sounds is known (for example, see Non-Patent Document 1). In this technique, lung sounds are separated based on a sparse expression that constitutes the lung sounds most simply by the sum of a Fourier transform signal and a wavelet signal. This sparse representation f (t) is classified as a breathing sound, and w (t) is classified as an intermittent rale.
特開２００４－３５７７５８号公報JP 2004-357758 A
　肺音を高速フーリエ変換およびその逆変換を用いて正常呼吸音と連続性ラ音に分別する方法は、周波数帯を適応的に選択したフィルタリングの一種である。このため、両者が同じ周波数成分を共有しているときには分別することができない。また、ここで正常呼吸音としている信号には、異常音である断続性ラ音が含まれている可能性がある。 The method of separating lung sounds into normal breath sounds and continuous rales using fast Fourier transform and its inverse transform is a kind of filtering in which frequency bands are adaptively selected. For this reason, when both share the same frequency component, it cannot be distinguished. In addition, the signal that is used as the normal breathing sound here may include an intermittent sound that is an abnormal sound.
　また、肺音信号をフーリエ変換信号とウェーブレット信号との和で最も簡潔に肺音を構成するスパース表現に基づき肺音を分離する手法では、元々の肺音信号に連続性ラ音が含まれている場合には、連続性ラ音がｆ（ｔ）とｗ（ｔ）のどちらか一方に分別されるとは限らない。このため、連続性ラ音の解析処理には対応できない。 Also, in the method of separating lung sounds based on the sparse representation that composes the lung sounds most simply by the sum of the Fourier transform signal and the wavelet signal, the original lung sound signal contains continuous rales. If there is, the continuous rarity is not always classified into either f (t) or w (t). For this reason, it cannot respond to the analysis process of continuous rales.
　このように従来技術では、多様な異常音を含む肺音信号を正確に分別することができない。 Thus, in the conventional technology, lung sound signals including various abnormal sounds cannot be accurately classified.
　そこで、本発明は、人間などの肺音から連続性ラ音および断続性ラ音をより正確に分別できるようにすることを目的とする。 Therefore, an object of the present invention is to enable more accurate separation of continuous and intermittent rales from lung sounds of humans and the like.
　上述の目的を達成するため、本発明は、生体音信号処理装置において、生体音信号の原音スペクトログラムを表現した原音行列をロバスト主成分分析でスパース行列と低ランク行列に分解するロバスト主成分分析部と、前記スパース行列を変換して前記生体音信号から連続性の生体音を得る連続性音処理部と、前記低ランク行列を変換して前記生体音信号から連続性の生体音を除外した生体音を得る非連続性音処理部と、を有することを特徴とする。 To achieve the above object, the present invention provides a robust principal component analysis unit that decomposes an original sound matrix representing an original sound spectrogram of a biological sound signal into a sparse matrix and a low rank matrix by robust principal component analysis in the biological sound signal processing apparatus. A continuous sound processing unit that converts the sparse matrix to obtain continuous biological sound from the biological sound signal, and a biological body that converts the low rank matrix and excludes continuous biological sound from the biological sound signal And a discontinuous sound processing unit that obtains sound.
　また、本発明は、生体音信号処理方法において、生体音の原音スペクトログラムを表現した原音行列をロバスト主成分分析でスパース行列と低ランク行列に分解するロバスト主成分分析工程と、前記スパース行列を変換して前記生体音信号から連続性の生体音を得る連続性音処理部を得る第２工程と、前記低ランク行列を変換して前記生体音信号から連続性の生体音を除外した生体音を得る第３工程と、を有することを特徴とする。 The present invention also relates to a biological sound signal processing method, a robust principal component analysis step of decomposing an original sound matrix representing an original sound spectrogram of a biological sound into a sparse matrix and a low rank matrix by robust principal component analysis, and converting the sparse matrix A second step of obtaining a continuous sound processing unit that obtains a continuous body sound from the body sound signal, and a body sound obtained by converting the low rank matrix and excluding the continuous body sound from the body sound signal. And obtaining a third step.
　また、本発明は、生体音信号処理プログラムにおいて、コンピュータを、生体音の原音スペクトログラムを表現した原音行列をロバスト主成分分析でスパース行列と低ランク行列に分解するロバスト主成分分析手段と、前記スパース行列を変換して前記生体音信号から連続性の生体音を得る連続性音処理手段と、前記低ランク行列を変換して前記生体音信号から連続性の生体音を除外した生体音を得る非連続性音処理手段として機能させることを特徴とする。 In the biological sound signal processing program, the present invention provides a robust principal component analysis means for decomposing an original sound matrix representing an original sound spectrogram of biological sound into a sparse matrix and a low rank matrix by robust principal component analysis, and the sparse A continuous sound processing means for converting a matrix to obtain continuous biological sounds from the biological sound signal; and a non-continuous sound processing means for obtaining biological sounds excluding continuous biological sounds from the biological sound signals by converting the low rank matrix It is characterized by functioning as a continuous sound processing means.
　本発明によれば、人間などの連続性ラ音および断続性ラ音をより正確に分別できる。 According to the present invention, it is possible to more accurately separate continuous and intermittent rales such as humans.
本発明に係る生体音信号処理装置の第１の実施の形態のブロック図である。1 is a block diagram of a first embodiment of a biological sound signal processing device according to the present invention. 本発明に係る生体音信号処理装置の第１の実施の形態における生体音信号処理方法のフローチャートである。It is a flowchart of the biological sound signal processing method in 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第１の実施の形態において処理を施す原音のグラフである。It is a graph of the original sound which processes in the 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第１の実施の形態において原音信号を短時間フーリエ変換して得られた原音スペクトログラムの値を要素に持つ行列の要素の値を濃淡で表した図である。It is the figure which represented the value of the element of the matrix which has the value of the original sound spectrogram obtained by carrying out the short-time Fourier transform of the original sound signal in the 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention by the light and shade. . 本発明に係る生体音信号処理装置の第１の実施の形態においてロバスト主成分分析で得た低ランク行列の要素の値を濃淡で表した図である。It is the figure which represented the value of the element of the low rank matrix obtained by the robust principal component analysis in 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention with the shading. 本発明に係る生体音信号処理装置の第１の実施の形態においてロバスト主成分分析で得たスパース行列の要素の値を濃淡で表した図である。It is the figure which represented the value of the element of the sparse matrix obtained by the robust principal component analysis in the 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention with the shading. 本発明に係る生体音信号処理装置の第１の実施の形態においてスパース行列を逆フーリエ変換して得られた肺音信号のグラフである。It is a graph of the lung sound signal obtained by carrying out the inverse Fourier transform of the sparse matrix in 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第１の実施の形態において低ランク行列を逆フーリエ変換して得られた肺音信号のグラフである。It is a graph of the lung sound signal obtained by carrying out the inverse Fourier transform of the low rank matrix in 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第１の実施の形態において低ランク行列を逆フーリエ変換して得られた肺音信号から抽出したフーリエ成分のグラフである。It is a graph of the Fourier component extracted from the lung sound signal obtained by carrying out the inverse Fourier transform of the low rank matrix in 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第１の実施の形態において低ランク行列を逆フーリエ変換して得られた肺音信号から抽出したウェーブレット成分のグラフである。It is a graph of the wavelet component extracted from the lung sound signal obtained by carrying out the inverse Fourier transform of the low rank matrix in 1st Embodiment of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第２の実施の形態のブロック図である。It is a block diagram of 2nd Embodiment of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第２の実施の形態生体音信号処理方法のフローチャートである。It is a flowchart of 2nd Embodiment biological sound signal processing method of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第３の実施の形態のブロック図である。It is a block diagram of 3rd Embodiment of the biological sound signal processing apparatus which concerns on this invention. 本発明に係る生体音信号処理装置の第３の実施の形態の生体音信号処理方法のフローチャートである。It is a flowchart of the biological sound signal processing method of 3rd Embodiment of the biological sound signal processing apparatus which concerns on this invention.
　本発明に係る生体音信号処理装置のいくつかの実施の形態を、図面を参照して説明する。なお、この実施の形態は単なる例示であり、本発明はこれに限定されない。同一または類似の構成には同一の符号を付し、重複する説明は省略する。 Several embodiments of the biological sound signal processing apparatus according to the present invention will be described with reference to the drawings. This embodiment is merely an example, and the present invention is not limited to this. The same or similar components are denoted by the same reference numerals, and redundant description is omitted.
［第１の実施の形態］
　図１は、本発明に係る生体音信号処理装置の第１の実施の形態のブロック図である。 [First Embodiment]
FIG. 1 is a block diagram of a first embodiment of a biological sound signal processing apparatus according to the present invention.
　生体音信号処理装置９０は、予備処理部１０とロバスト主成分分析部４０とスパース行列格納部２２と低ランク行列格納部３２と連続性音処理部２０と非連続性音処理部３０とを有している。生体音信号処理装置９０は、たとえば１台のコンピュータ上に構築される。生体音信号処理装置９０は、ネットワークで結合された複数台のコンピュータ上に構築されていてもよい。 The biological sound signal processing device 90 includes a preliminary processing unit 10, a robust principal component analysis unit 40, a sparse matrix storage unit 22, a low rank matrix storage unit 32, a continuous sound processing unit 20, and a discontinuous sound processing unit 30. is doing. The biological sound signal processing device 90 is constructed on, for example, one computer. The biological sound signal processing device 90 may be constructed on a plurality of computers connected by a network.
　予備処理部１０は、入力部１２とフーリエ変換部１４と行列生成部１６とを有している。連続性音処理部２０は、連続性音スペクトログラム生成部２４と第１逆フーリエ変換部２６とを有している。非連続性音処理部３０は、非連続性音スペクトログラム生成部３４と第２逆フーリエ変換部３６と信号抽出部３８とを有している。 The preliminary processing unit 10 includes an input unit 12, a Fourier transform unit 14, and a matrix generation unit 16. The continuous sound processing unit 20 includes a continuous sound spectrogram generation unit 24 and a first inverse Fourier transform unit 26. The discontinuous sound processing unit 30 includes a discontinuous sound spectrogram generation unit 34, a second inverse Fourier transform unit 36, and a signal extraction unit 38.
　予備処理部１０は、生体音の原音スペクトログラムを表現した原音行列を生成する。生体音は、図示しない電子聴診器のような生体音検出装置で検出し、電気信号として生体音信号処理装置９０に与えられる。 The preliminary processing unit 10 generates an original sound matrix that represents an original sound spectrogram of a biological sound. The biological sound is detected by a biological sound detection device such as an electronic stethoscope (not shown), and is supplied to the biological sound signal processing device 90 as an electrical signal.
　ロバスト主成分分析部４０は、予備処理部１０が生成した原音行列をロバスト主成分分析し、スパース行列と低ランク行列を得る。スパース行列は、スパース行列格納部２２に格納される。低ランク行列は、低ランク行列格納部３２に格納される。 The robust principal component analysis unit 40 performs a robust principal component analysis on the original sound matrix generated by the preliminary processing unit 10 to obtain a sparse matrix and a low rank matrix. The sparse matrix is stored in the sparse matrix storage unit 22. The low rank matrix is stored in the low rank matrix storage unit 32.
　連続性音処理部２０は、スパース行列格納部２２に格納されたスパース行列を処理して、原音中の連続性音を生成する。非連続性音処理部３０は、低ランク行列格納部３２に格納された低ランク行列を処理して、原音中の非連続性音を生成する。 The continuity sound processing unit 20 processes the sparse matrix stored in the sparse matrix storage unit 22 to generate a continuity sound in the original sound. The discontinuous sound processing unit 30 processes the low rank matrix stored in the low rank matrix storage unit 32 to generate a discontinuous sound in the original sound.
　連続性音処理部２０および非連続性音処理部３０が生成した連続性音および非連続性音は、たとえばＤ／Ａ変換を施されて図示しないスピーカなどによって出力される。連続性音および非連続性音の波形をディスプレイに表示してもよい。あるいは、連続性音および非連続性音の信号を外部の装置に送信し、その外部装置で異常検知などを行ってもよい。 The continuous sound and the discontinuous sound generated by the continuous sound processing unit 20 and the discontinuous sound processing unit 30 are, for example, subjected to D / A conversion and output by a speaker (not shown). The waveform of continuous sound and discontinuous sound may be displayed on the display. Alternatively, a continuity sound signal and a discontinuity sound signal may be transmitted to an external device, and abnormality detection or the like may be performed by the external device.
　次に、この生体音信号処理装置９０を用いた生体音信号処理方法を説明する。 Next, a biological sound signal processing method using this biological sound signal processing apparatus 90 will be described.
　図２は、本実施の形態における生体音信号処理方法のフローチャートである。 FIG. 2 is a flowchart of the biological sound signal processing method in the present embodiment.
　まず、入力部１２が図示しない電子聴診器のような生体音検出装置からの生体音を検出した信号を取りこむ。検出された生体音を原音と呼ぶこととする。原音をたとえば電気的に表現した信号を原音信号ｓ（ｔ）と呼ぶこととする。生体音とは、たとえば人間の肺音である。入力部１２は、原音信号がアナログ信号の場合は、Ａ／Ｄ変換して原音信号をデジタルデータに変換する。 First, the input unit 12 captures a signal that detects a body sound from a body sound detection device such as an electronic stethoscope (not shown). The detected biological sound is called an original sound. For example, a signal that electrically represents the original sound is referred to as an original sound signal s (t). The biological sound is, for example, a human lung sound. When the original sound signal is an analog signal, the input unit 12 performs A / D conversion to convert the original sound signal into digital data.
　図３は、本実施の形態において処理を施す肺音信号のグラフである。図３において、横軸は経過時間（秒）、縦軸は信号強度（振幅）を示す。 FIG. 3 is a graph of lung sound signals to be processed in the present embodiment. In FIG. 3, the horizontal axis represents elapsed time (seconds), and the vertical axis represents signal intensity (amplitude).
　本実施の形態では、非特許文献２の付録のＣＤの第６０トラックに記録された肺音を原音として処理を行う。 In this embodiment, the lung sound recorded on the 60th track of the CD in the appendix of Non-Patent Document 2 is used as the original sound.
　次に、フーリエ変換部１４が原音信号ｓ（ｔ）を短時間フーリエ変換し、時間周波数領域の複素量で表される複素サウンドスペクトログラム（以下、スペクトログラムと称する。）を得る（ステップ１）。 Next, the Fourier transform unit 14 performs a short-time Fourier transform on the original sound signal s (t) to obtain a complex sound spectrogram (hereinafter referred to as a spectrogram) represented by a complex quantity in the time-frequency domain (step 1).
　より具体的には、原音信号ｓ（ｔ）に時間窓関数をずらしながら掛けたものを離散フーリエ変換して、原音スペクトログラムＳ（ω，ｔ）を得る。原音スペクトログラムＳ（ω，ｔ）は、複素数の値を持ち、時間窓関数の位置を表す時刻ｔにおいて信号を構成している角周波数ωの成分の振幅と位相を表す。 More specifically, the original sound spectrogram S (ω, t) is obtained by subjecting the original sound signal s (t) multiplied by the time window function to a discrete Fourier transform. The original sound spectrogram S (ω, t) has a complex value and represents the amplitude and phase of the component of the angular frequency ω constituting the signal at time t representing the position of the time window function.
　時刻ｔは、時間窓関数をずらす時間幅Δｔ２で離散値をとる。この時間窓関数をずらす時間幅Δｔ２は、時間窓幅Δｔ１を超えないものとする。すなわち、Δｔ２＜Δｔ１である。また、角周波数ωは時間窓幅Δｔ１の逆数に比例する間隔で離散化されている。 Time t takes a discrete value with a time width Δt2 for shifting the time window function. It is assumed that the time width Δt2 for shifting the time window function does not exceed the time window width Δt1. That is, Δt2 <Δt1. Further, the angular frequency ω is discretized at an interval proportional to the reciprocal of the time window width Δt1.
　その後、行列生成部１６が原音スペクトログラムＳ（ω，ｔ）の振幅｜Ｓ（ω，ｔ）｜を要素に持つ原音行列Ｄを作成する（ステップ２）。原音行列Ｄの行番号ｉと列番号ｊは、ｉ番目の角周波数ωｉ、ｊ番目の時刻ｔｊに対応するものとする。行列Ｄの第ｉ行ｊ列の要素Ｄｉｊは、原音スペクトログラムＳ（ω，ｔ）を構成する複素数Ｓ（ωｉ，ｔｊ）の絶対値とする。 Thereafter, the matrix generation unit 16 creates an original sound matrix D having the amplitude | S (ω, t) | of the original sound spectrogram S (ω, t) as an element (step 2). The row number i and the column number j of the original sound matrix D correspond to the i-th angular frequency ω i and the j-th time t j . An element Dij in the i-th row and j-th column of the matrix D is an absolute value of a complex number S (ω i , t j ) constituting the original sound spectrogram S (ω, t).
　図４は、本実施の形態において原音信号を短時間フーリエ変換して得られた原音スペクトログラムの値を要素に持つ原音行列の要素の値を濃淡で表した図である。 FIG. 4 is a diagram in which the values of the elements of the original sound matrix having the values of the original sound spectrogram obtained by performing a short-time Fourier transform on the original sound signal in this embodiment are represented by shading.
　ステップ３では、ロバスト主成分分析部４０が原音行列Ｄを低ランク行列Ａとスパース行列Ｅの和の形になるように分解する。ここでは、非特許文献３に倣って、このような行列の分解をロバスト主成分分析と呼ぶ。ある行列を低ランク行列とスパース行列の和の形になるように分解する算法としては、たとえば非特許文献４に、拡張ラグランジュ法を改良した収束の早い算法が提案されている。 In step 3, the robust principal component analysis unit 40 decomposes the original sound matrix D so as to be the sum of the low rank matrix A and the sparse matrix E. Here, in accordance with Non-Patent Document 3, such matrix decomposition is called robust principal component analysis. As an algorithm for decomposing a matrix so as to be in the form of the sum of a low rank matrix and a sparse matrix, for example, Non-Patent Document 4 proposes a fast convergence algorithm improved from the extended Lagrangian method.
　通常の主成分分析では、与えられた行列を低ランク行列で近似する。その低ランク行列は、与えられた行列の主要な固有値（または特異値）のみに付随する固有ベクトルの積によって構成される。 In normal principal component analysis, a given matrix is approximated by a low rank matrix. The low rank matrix is constructed by the product of eigenvectors associated with only the main eigenvalues (or singular values) of a given matrix.
　一方、本実施の形態で用いるロバスト主成分分析では、与えられた原音行列Ｄを、その要素の一部にのみ修正を許しながら低ランク行列Ａで近似する。原音行列Ｄは低ランク行列Ａと修正量を表すスパース行列Ｅの和に分解される。その際、行列Ａのランク（階数）と、修正する要素の数（行列Ｅの非ゼロ要素の数）は、可能な限り共に小さいものとする。この場合の解の一意性は非特許文献３に開示されている。 On the other hand, in the robust principal component analysis used in the present embodiment, a given original sound matrix D is approximated by a low rank matrix A while allowing only a part of its elements to be modified. The original sound matrix D is decomposed into a sum of a low rank matrix A and a sparse matrix E representing a correction amount. At that time, the rank (rank) of the matrix A and the number of elements to be corrected (number of non-zero elements of the matrix E) are both as small as possible. The uniqueness of the solution in this case is disclosed in Non-Patent Document 3.
　図５は、本実施の形態においてロバスト主成分分析で得た低ランク行列の要素の値を濃淡で表した図である。図６は、本実施の形態においてロバスト主成分分析で得たスパース行列の要素の値を濃淡で表した図である。 FIG. 5 is a diagram in which the values of the elements of the low rank matrix obtained by the robust principal component analysis in the present embodiment are represented by shading. FIG. 6 is a diagram in which the values of the elements of the sparse matrix obtained by the robust principal component analysis in the present embodiment are represented by shading.
　低ランク行列Ａは、主要な固有ベクトルの積によって構成し易い行列Ｄの成分を要素に持つ。ゆえに、低ランク行列Ａの行または列が呈する図５の模様は、原音行列の要素の値を濃淡で表した図４に見られる縦縞・横縞模様のように、類似の模様が複数現れる傾向がある。 The low rank matrix A has elements of a matrix D that can be easily constructed by the product of main eigenvectors. Therefore, the pattern in FIG. 5 exhibited by the rows or columns of the low rank matrix A tends to appear in a plurality of similar patterns, such as the vertical stripes and horizontal stripe patterns shown in FIG. 4 in which the values of the elements of the original sound matrix are represented by shading. is there.
　一方、そのような低ランク行列Ａによって原音行列Ｄを近似するために除外された成分をスパース行列Ｅは要素として持つ。ゆえに、図６に示されるように、スパース行列Ｅは、縦縞・横縞などの規則性を持たない、任意の曲線状または斑点状の模様を呈する。 On the other hand, the sparse matrix E has the components excluded to approximate the original sound matrix D by such a low rank matrix A as elements. Therefore, as shown in FIG. 6, the sparse matrix E exhibits an arbitrary curved or spotted pattern having no regularity such as vertical stripes and horizontal stripes.
　ステップ４では、ステップ３で得られたスパース行列Ｅに対応する連続性音スペクトログラムＥ（ω，ｔ）を生成する。連続性音スペクトログラムＥ（ω，ｔ）は、連続性音スペクトログラム生成部２４がスパース行列格納部２２からスパース行列Ｅを読み込んで生成する。連続性音スペクトログラムＥ（ω，ｔ）を構成する複素数Ｅ（ωｉ，ｔｊ）は、前記スパース行列の要素を振幅とし、前記原音スペクトログラの偏角を偏角とする複素数とする。すなわち複素数Ｅ（ωｉ，ｔｊ）は、行列Ｅの第ｉ行ｊ列の要素Ｅｉｊと原音スペクトログラムＳ（ω，ｔ）を構成する複素数Ｓ（ωｉ，ｔｊ）の偏角θｉｊから次式によって得られる。 In step 4, a continuous sound spectrogram E (ω, t) corresponding to the sparse matrix E obtained in step 3 is generated. The continuity sound spectrogram E (ω, t) is generated by the continuity sound spectrogram generation unit 24 reading the sparse matrix E from the sparse matrix storage unit 22. The complex number E (ω i , t j ) constituting the continuity sound spectrogram E (ω, t) is a complex number having an element of the sparse matrix as an amplitude and an argument of the original sound spectrogram as an argument. That is, the complex number E (ω i , t j ) is the argument θ ij of the complex number S (ω i , t j ) that constitutes the element E ij in the i-th row and j-th column of the matrix E and the original sound spectrogram S (ω, t). Is obtained by the following equation.
　Ｅ（ωｉ，ｔｊ）＝Ｅｉｊ　（cos(θｉｊ)＋　ｉsin(θｉｊ)） E (ω i , t j ) = E ij (cos (θ ij ) + isin (θ ij ))
　ステップ５では、第１逆フーリエ変換部２６が連続性音スペクトログラムＥ（ω，ｔ）を短時間逆フーリエ変換し、肺音信号ｅ（ｔ）を得る。より具体的には、連続性音スペクトログラムＥ（ω，ｔ）の各時刻ｔｊ毎に複素スペクトルＥ（ω，ｔｊ）を短時間逆フーリエ変換し、各時刻ｔｊにおける時間窓関数内の肺音信号を得る。時間窓関数の重なりに応じて肺音信号を平均化することで肺音信号ｅ（ｔ）を得る。 In step 5, the first inverse Fourier transform unit 26 performs a short-time inverse Fourier transform on the continuous sound spectrogram E (ω, t) to obtain a lung sound signal e (t). More specifically, continuity sound spectrogram E (omega, t) complex spectrum E (omega, t j) for each time t j of the inverse Fourier transform short time, in the time window function at each time t j A lung sound signal is obtained. The lung sound signal e (t) is obtained by averaging the lung sound signals according to the overlap of the time window functions.
　図７は、本実施の形態においてスパース行列を逆フーリエ変換して得られた肺音信号のｅ（ｔ）グラフである。 FIG. 7 is an e (t) graph of a lung sound signal obtained by performing an inverse Fourier transform on a sparse matrix in the present embodiment.
　図７は、デジタルデータとして第１逆フーリエ変換部２６が出力した肺音信号ｅ（ｔ）をアナログ信号のように模擬して表示したものである。肺音信号ｅ（ｔ）を図示しない生体音出力部でＤ／Ａ変換し、図示しないスピーカなどによって肺音として再現すると、連続性ラ音となる。 FIG. 7 shows the lung sound signal e (t) output from the first inverse Fourier transform unit 26 as digital data, which is simulated and displayed like an analog signal. When the lung sound signal e (t) is D / A converted by a biological sound output unit (not shown) and reproduced as a lung sound by a speaker (not shown) or the like, it becomes a continuous sound.
　本実施の形態において、処理の入力として使用した原音信号ｓ（ｔ）は、非特許文献２に付録のＣＤの第６０トラックから採用した連続性ラ音（高音性の笛音）を含む肺音である。この連続性ラ音は、原音信号ｓ（ｔ）のスペクトログラムを表す図４において曲線状の模様を呈しており、図６に示したスパース行列Ｅが表すスペクトログラムへ明確に分離されている。このことから、上記のステップ１からステップ５の処理により、連続性ラ音がステップ５の肺音信号ｅ（ｔ）として良好に分離されていることが分かる。 In the present embodiment, the original sound signal s (t) used as an input of processing is a lung sound including a continuous ra sound (high-pitched whistle sound) adopted from Non-Patent Document 2 from the 60th track of the CD of the appendix. It is. The continuous ra sound has a curved pattern in FIG. 4 representing the spectrogram of the original sound signal s (t), and is clearly separated into the spectrogram represented by the sparse matrix E shown in FIG. From this, it can be seen that the continuous sound is well separated as the lung sound signal e (t) in step 5 by the processing from step 1 to step 5 described above.
　この例で使用した連続性ラ音は、高音性の笛音であるが、低音性のいびき音についても分離することが可能である。本実施例と同様に、連続性ラ音以外の音（呼吸音や断続性ラ音）のスペクトログラムが低ランク行列Ａによって表され易ければ、ステップ３のロバスト主成分分析において、低ランク行列Ａで行列Ｄを近似するために除外された成分（スパース行列Ｅ）として連続性ラ音のスペクトログラムが分離される。 The continuous rales used in this example are high-pitched whistle sounds, but low-pitched snoring sounds can also be separated. Similar to the present embodiment, if the spectrogram of sounds other than continuous rales (breathing sounds and intermittent rachunes) is easily represented by the low rank matrix A, the low rank matrix A in the robust principal component analysis of step 3 is performed. The spectrogram of the continuous ra-tone is separated as a component (sparse matrix E) excluded to approximate the matrix D.
　ステップ６では、ステップ３にて分離した低ランク行列Ａに対応する非連続性音スペクトログラムＡ（ω，ｔ）を非連続性音スペクトログラム生成部３４が生成する。生成の方法はステップ４と同様である。すなわち、非連続性音スペクトログラムＡ（ω，ｔ）を構成する複素数Ａ（ωｉ，ｔｊ）は、前記低ランク行列の要素を振幅とし、前記原音スペクトログラの偏角を偏角とする複素数とし、行列Ａの第ｉ行ｊ列の要素Ａｉｊと原音スペクトログラムＳ（ω，ｔ）を構成する複素数Ｓ（ωｉ，ｔｊ）の偏角θｉｊから次式によって得られる。 In step 6, the discontinuous sound spectrogram generation unit 34 generates the discontinuous sound spectrogram A (ω, t) corresponding to the low rank matrix A separated in step 3. The generation method is the same as in Step 4. That is, the complex number A (ω i , t j ) constituting the discontinuous sound spectrogram A (ω, t) is a complex number having an element of the low rank matrix as an amplitude and an argument of the original sound spectrogram as an argument. And the element A ij in the i-th row and j-th column of the matrix A and the argument θ ij of the complex number S (ω i , t j ) constituting the original sound spectrogram S (ω, t), is obtained by the following equation.
　Ａ（ωｉ，ｔｊ）＝Ａｉｊ　（cos(θｉｊ)＋　ｉsin(θｉｊ)） A (ω i , t j ) = A ij (cos (θ ij ) + isin (θ ij ))
　ステップ７では、第２逆フーリエ変換部３６が非連続性音スペクトログラムＡ（ω，ｔ）を短時間逆フーリエ変換し、肺音信号ａ（ｔ）を得る。この方法は、ステップ５と同様である。すなわち、非連続性音スペクトログラムＡ（ω，ｔ）の各時刻ｔｊ毎に複素スペクトルＡ（ω，ｔｊ）を短時間逆フーリエ変換し、各時刻ｔｊにおける時間窓関数内の肺音信号を得る。窓関数の重なりに応じて肺音信号を平均化することで肺音信号ａ（ｔ）を得る。 In step 7, the second inverse Fourier transform unit 36 performs a short-time inverse Fourier transform on the discontinuous sound spectrogram A (ω, t) to obtain a lung sound signal a (t). This method is the same as step 5. That is, discontinuous sound spectrogram A (omega, t) complex spectrum A (omega, t j) for each time t j of the inverse Fourier transform short time, lung sounds signals within a time window function at each time t j Get. A lung sound signal a (t) is obtained by averaging the lung sound signals according to the overlap of the window functions.
　図８は、本実施の形態において低ランク行列Ａを逆フーリエ変換して得られた肺音信号ａ（ｔ）のグラフである。 FIG. 8 is a graph of the lung sound signal a (t) obtained by performing inverse Fourier transform on the low rank matrix A in the present embodiment.
　図８は、デジタルデータとして第２逆フーリエ変換部３６が出力した肺音信号をアナログ信号のように模擬して表示したものである。肺音信号ａ（ｔ）を図示しない生体音出力部でＤ／Ａ変換し、図示しないスピーカなどによって肺音として再現すると、原音信号ｓ（ｔ）から連続性音を除去した音となる。 FIG. 8 shows the lung sound signal output from the second inverse Fourier transform unit 36 as digital data by simulating and displaying it as an analog signal. When the lung sound signal a (t) is D / A converted by a biological sound output unit (not shown) and reproduced as a lung sound by a speaker or the like (not shown), a sound obtained by removing the continuous sound from the original sound signal s (t) is obtained.
　ステップ８では、更に非特許文献１の手法で、信号抽出部３８が肺音信号ｅ（ｔ）から肺音信号ｆ（ｔ）と肺音信号ｗ（ｔ）を抽出する。非特許文献１の手法は、スパース表現に基づき肺音を抽出するものである。ここでは、肺音信号ａ（ｔ）を、フーリエ合成信号（ｆ（ｔ））とウェーブレット合成信号（ｗ（ｔ））との和で表現する。その際、非ゼロのフーリエ成分の数と非ゼロのウェーブレット成分の数は、可能な限り共に小さいものとする。このための算法としては、たとえば非特許文献１に開示されている。 In step 8, the signal extraction unit 38 further extracts the lung sound signal f (t) and the lung sound signal w (t) from the lung sound signal e (t) by the method of Non-Patent Document 1. The method of Non-Patent Document 1 extracts lung sounds based on sparse representation. Here, the lung sound signal a (t) is expressed by the sum of the Fourier synthesized signal (f (t)) and the wavelet synthesized signal (w (t)). At this time, the number of non-zero Fourier components and the number of non-zero wavelet components are both as small as possible. For example, Non-Patent Document 1 discloses an algorithm for this purpose.
　図９は、本実施の形態において肺音信号ｅ（ｔ）から抽出したフーリエ成分である肺音信号ｆ（ｔ）のグラフである。 FIG. 9 is a graph of a lung sound signal f (t) that is a Fourier component extracted from the lung sound signal e (t) in the present embodiment.
　図９は、デジタルデータとして信号抽出部３８が出力した肺音信号ｆ（ｔ）をアナログ信号のように模擬して表示したものである。肺音信号ｆ（ｔ）を図示しない生体音出力部でＤ／Ａ変換し、図示しないスピーカなどによって肺音として再現すると、原音信号ｓ（ｔ）中の呼吸音が出力される。 FIG. 9 shows the lung sound signal f (t) output from the signal extraction unit 38 as digital data by simulating it as an analog signal. When the lung sound signal f (t) is D / A converted by a biological sound output unit (not shown) and reproduced as a lung sound by a speaker (not shown), a respiratory sound in the original sound signal s (t) is output.
　図１０は、本実施の形態において肺音信号ｅ（ｔ）から抽出したウェーブレット成分である肺音信号ｗ（ｔ）のグラフである。 FIG. 10 is a graph of the lung sound signal w (t), which is a wavelet component extracted from the lung sound signal e (t) in the present embodiment.
　図１０は、デジタルデータとして信号抽出部３８が出力した肺音信号ｗ（ｔ）をアナログ信号のように模擬して表示したものである。肺音信号ｗ（ｔ）を図示しない生体音出力部でＤ／Ａ変換し、図示しないスピーカなどによって肺音として再現すると、原音信号ｓ（ｔ）中の断続性ラ音が出力される。 FIG. 10 shows the lung sound signal w (t) output from the signal extraction unit 38 as digital data by simulating it as an analog signal. When the lung sound signal w (t) is D / A converted by a biological sound output unit (not shown) and reproduced as a lung sound by a speaker (not shown) or the like, an intermittent ra sound in the original sound signal s (t) is output.
　本実施の形態において、原音信号ｓ（ｔ）は、非特許文献２付録のＣＤに第６０トラックから採用した連続性ラ音データ（高音性の笛音）である。しかし、図３に示した原音信号ｓ（ｔ）の波形から、連続性ラ音以外にも、断続性の音が混入していることを確認できる。また、図４に示した原音信号ｓ（ｔ）を表すスペクトログラムから、曲線状の模様を呈する連続性ラ音以外にも、縦縞模様を呈する断続性の音（断続性ラ音）や、低周波帯に継続する音（呼吸音）の存在を確認できる。連続性ラ音以外のこれらの音は、図５に示した低ランク行列Ａが表すスペクトログラムへ明確に分離されている。このことから、上記のステップ１からステップ３およびステップ６からステップ８の処理により、呼吸音と断続性ラ音が良好に抽出されていることが分かる。 In the present embodiment, the original sound signal s (t) is continuous sound data (high-pitched whistle sound) adopted from the 60th track in the CD of Appendix of Non-Patent Document 2. However, it can be confirmed from the waveform of the original sound signal s (t) shown in FIG. 3 that intermittent sound is mixed in addition to continuous rarity. Further, from the spectrogram representing the original sound signal s (t) shown in FIG. 4, in addition to the continuous ra sound that exhibits a curvilinear pattern, an intermittent sound that exhibits a vertical stripe pattern (intermittent ra sound), a low frequency The presence of sound (breathing sound) that continues in the belt can be confirmed. These sounds other than the continuous rales are clearly separated into spectrograms represented by the low rank matrix A shown in FIG. From this, it can be seen that the breathing sound and the intermittent rar sound are extracted satisfactorily by the processing from Step 1 to Step 3 and Step 6 to Step 8.
　このように、本実施の形態によれば、人間などの肺音から呼吸音、連続性ラ音、断続性ラ音をより正確に分別できる。 Thus, according to this embodiment, it is possible to more accurately classify respiratory sounds, continuous rales, and intermittent rales from lung sounds of humans and the like.
　本実施の形態では、人間の肺音から呼吸音、連続性ラ音、断続性ラ音を分離・抽出したが、人間以外の動物であっても、その動物が発する生体音に呼吸音、連続性ラ音および断続性ラ音に類する音が含まれる場合には、本実施の形態と同様にそれらの音を分離・抽出することができる。 In the present embodiment, respiratory sounds, continuous rales, and intermittent rales are separated and extracted from human lung sounds. In the case where sounds similar to a natural ra sound and an intermittent ra sound are included, those sounds can be separated and extracted as in the present embodiment.
［第２の実施の形態］
　図１１は、本発明に係る生体音信号処理装置の第２の実施の形態のブロック図である。図１２は、本実施の形態における生体音信号処理方法のフローチャートである。 [Second Embodiment]
FIG. 11 is a block diagram of a second embodiment of the biological sound signal processing device according to the present invention. FIG. 12 is a flowchart of the biological sound signal processing method in the present embodiment.
　本実施の形態は、第１の実施の形態と、プロセスの全体として複素行列を用いる点が異なる。このため、第１の実施の形態における行列生成部１６は、本実施の形態の生体音信号処理装置には存在しない。 This embodiment is different from the first embodiment in that a complex matrix is used as a whole process. For this reason, the matrix generation part 16 in 1st Embodiment does not exist in the biological sound signal processing apparatus of this Embodiment.
　本実施の形態では、第１の実施の形態におけるステップ２を省略し、ステップ３においてロバスト主成分分析部４０は原音スペクトログラムＳ（ω，ｔ）を直接ロバスト主成分分析して低ランク行列Ａとスパース行列Ｅを得る。分離した肺音信号を得るステップ５と７、およびステップ８以降は、第１の実施の形態と同じである。 In the present embodiment, step 2 in the first embodiment is omitted, and in step 3, the robust principal component analysis unit 40 directly performs a robust principal component analysis on the original sound spectrogram S (ω, t) to obtain a low rank matrix A and Get the sparse matrix E. Steps 5 and 7 for obtaining a separated lung sound signal, and step 8 and subsequent steps are the same as those in the first embodiment.
　第１の実施の形態における複素行列を実数行列に変換し、ロバスト主成分分析の後に、複素行列に再度変換するというステップを、本実施の形態では削除している。振幅と位相を表す複素数の行列を分離するので、振幅のみを分離する実施形態１よりも分離の性能向上を期待できる。なお、ロバスト主成分分析の対象の行列が複素行列であるため、計算時間が第１の実施の形態よりも若干長くなるものの、適切な解法を用いれば、十分、実用に耐える。 In the present embodiment, the step of converting the complex matrix in the first embodiment into a real matrix and converting it again into a complex matrix after robust principal component analysis is omitted in this embodiment. Since the matrix of complex numbers representing the amplitude and the phase is separated, the separation performance can be improved as compared with the first embodiment in which only the amplitude is separated. Note that, since the matrix to be subjected to the robust principal component analysis is a complex matrix, the calculation time is slightly longer than that of the first embodiment. However, if an appropriate solution is used, it is sufficiently practical.
［第３の実施の形態］
　図１３は、本発明に係る生体音信号処理装置の第３の実施の形態のブロック図である。図１４は、本実施の形態における生体音信号処理方法のフローチャートである。 [Third Embodiment]
FIG. 13 is a block diagram of a third embodiment of the biological sound signal processing device according to the present invention. FIG. 14 is a flowchart of the biological sound signal processing method in the present embodiment.
　本実施の形態は、第１の実施の形態と、短時間フーリエ変換および短時間逆フーリエ変換の代わりに、短時間コサイン変換および短時間逆コサイン変換を用いる点が異なる。本実施の形態の生体音信号処理装置では、第１の実施の形態における短時間フーリエ変換部１４の代わりに、短時間コサイン変換部４１が設けられている。 This embodiment is different from the first embodiment in that short-time cosine transform and short-time inverse cosine transform are used instead of short-time Fourier transform and short-time inverse Fourier transform. In the biological sound signal processing device of the present embodiment, a short-time cosine transform unit 41 is provided instead of the short-time Fourier transform unit 14 in the first embodiment.
　本実施の形態は、第１の実施の形態と、プロセスの全体として実数行列を用いる点が異なる。このため、第１の実施の形態における行列生成部１６は、本実施の形態の生体音信号処理装置には存在しない。 This embodiment is different from the first embodiment in that a real number matrix is used as a whole process. For this reason, the matrix generation part 16 in 1st Embodiment does not exist in the biological sound signal processing apparatus of this Embodiment.
　本実施の形態では、第１の実施の形態におけるステップ２を省略し、ステップ３においてロバスト主成分分析部５０は短時間コサイン変換部で得た実数の原音スペクトログラムＳ（ω，ｔ）を直接ロバスト主成分分析して低ランク行列Ａとスパース行列Ｅを得る。また、本実施の形態の生体音信号処理装置では、第１の実施の形態における第１短時間逆フーリエ変換部２６の代わりに、短時間逆コサイン変換部５２が設けられている。本実施の形態の生体音信号処理装置では、第１の実施の形態における第２短時間逆フーリエ変換部３６の代わりに、短時間逆コサイン変換部５３が設けられている。 In the present embodiment, step 2 in the first embodiment is omitted, and in step 3, the robust principal component analysis unit 50 directly robusts the real original spectrogram S (ω, t) obtained by the short-time cosine transform unit. Principal component analysis is performed to obtain a low rank matrix A and a sparse matrix E. Further, in the biological sound signal processing device of the present embodiment, a short-time inverse cosine transform unit 52 is provided instead of the first short-time inverse Fourier transform unit 26 in the first embodiment. In the biological sound signal processing apparatus according to the present embodiment, a short-time inverse cosine transform unit 53 is provided instead of the second short-time inverse Fourier transform unit 36 in the first embodiment.
　コサイン変換は実信号のフーリエ変換に対して半分の記憶領域で処理できる利点がある。ただし、信号が偶関数であることを仮定した処理なので、他の実施形態より分離の性能がやや劣る可能性がある。 The cosine transform has the advantage that it can be processed in half the storage area compared to the Fourier transform of the real signal. However, since the process assumes that the signal is an even function, the separation performance may be slightly inferior to that of the other embodiments.
第１，第２，第３のいずれの実施の形態においても、低ランク行列Ａからは、特異値分解によって
　Ａ＝ＵＫＶＴ
となる行列Ｕ、行列Ｋ、行列Ｖが求められる。低ランク行列Ａのサイズがｍ×ｎ、ランク（階数）がｒであるとすると、行列Ｋは、特異値を対角にもつ対角のｒ次正方行列であり、行列ＵおよびＶはそれぞれｒ本の左特異ベクトルと右特異ベクトルからなるｍ×ｒ行列、ｎ×ｒ行列である。なお、ＶＴは、行列Ｖが実行列の場合はＶの転置行列であり、行列Ｖが複素行列の場合はＶの共役転置行列である。 In any of the first, second, and third embodiments, A = UKV T is obtained from the low rank matrix A by singular value decomposition.
A matrix U, a matrix K, and a matrix V are obtained. If the size of the low-rank matrix A is m × n and the rank (rank) is r, the matrix K is a diagonal r-order square matrix with singular values on the diagonal, and the matrices U and V are r An m × r matrix and an n × r matrix composed of a left singular vector and a right singular vector Note that V T is a transposed matrix of V when the matrix V is a real matrix, and a conjugate transposed matrix of V when the matrix V is a complex matrix.
左特異ベクトルは、行列Ａの列ベクトルを合成できる基底である。低ランク行列Ａの列ベクトルは、連続性の生体音以外の音、特に呼吸音と断続音の瞬時周波数スペクトルを表している。ゆえに、左特異ベクトルは、これらの音の瞬時周波数スペクトルを構成するための基底である。また、右特異ベクトルは、任意の時刻における瞬時周波数スペクトルの内訳を示している。すなわち、ｊ番目の右特異ベクトルの成分は、ｊ番目の左特異ベクトルの基底がそれぞれどの時刻でどの程度現れるかを表している。 The left singular vector is a basis on which the column vector of the matrix A can be synthesized. The column vector of the low rank matrix A represents the instantaneous frequency spectrum of sounds other than continuous biological sounds, particularly breathing sounds and intermittent sounds. Thus, the left singular vector is the basis for constructing the instantaneous frequency spectrum of these sounds. The right singular vector indicates the breakdown of the instantaneous frequency spectrum at an arbitrary time. That is, the component of the j-th right singular vector represents how much the base of the j-th left singular vector appears at each time.
したがって、低ランク行列Ａの特異値分解によって、瞬時周波数スペクトルの基底からなる行列Ｕ、その内訳を表すＶが得られる。同じ種類の音は類似した瞬時周波数スペクトルの内訳を持つことから、左右特異ベクトルを利用して同種の呼吸音や断続音を判別する分類に応用できる。 Therefore, by the singular value decomposition of the low rank matrix A, a matrix U composed of the basis of the instantaneous frequency spectrum and V representing the breakdown are obtained. Since the same type of sound has a similar breakdown of the instantaneous frequency spectrum, it can be applied to classification for discriminating the same type of breathing sound and intermittent sound using the left and right singular vectors.
　なお、特異値分解は、行列を構成する行ベクトルの集合および列ベクトルの集合について、主成分を求める行列分解の一手法である。特異値分解によって、主成分の大きさを表す特異値および主成分の向きを表す正規直交基底が得られる。ただし、ゼロの特異値に付随する特異ベクトルは、一意に定まらない。また、対応する左右の特異ベクトルは、符号または大きさ１の複素数倍の任意性がある。しかし、本実施の形態で得られる低ランク行列Ａは、非ゼロの特異値と、それらに付随する左右特異ベクトルによって構成されているため、一意に定まる。 Note that singular value decomposition is a method of matrix decomposition for obtaining principal components for a set of row vectors and a set of column vectors constituting a matrix. By singular value decomposition, a singular value representing the size of the principal component and an orthonormal basis representing the direction of the principal component are obtained. However, the singular vector associated with a singular value of zero is not uniquely determined. Corresponding singular vectors on the left and right are arbitrary in terms of a code or a complex multiple of size 1. However, the low rank matrix A obtained in the present embodiment is uniquely determined because it is composed of non-zero singular values and the left and right singular vectors associated therewith.
１０…予備処理部、１２…入力部、１４…フーリエ変換部、１６…行列生成部、２０…連続性音処理部、２２…スパース行列格納部、２４…連続性音スペクトログラム生成部、２６…第１逆フーリエ変換部、３０…非連続性音処理部、３２…低ランク行列格納部、３４…非連続性音スペクトログラム生成部、３６…第２逆フーリエ変換部、３８…信号抽出部、４０…ロバスト主成分分析部、４１…コサイン変換部、５０…ロバスト主成分分析部、５２…逆コサイン変換部、５３…逆コサイン変換部、９０…生体音信号処理装置
  DESCRIPTION OF SYMBOLS 10 ... Preliminary processing part, 12 ... Input part, 14 ... Fourier transform part, 16 ... Matrix generation part, 20 ... Continuous sound processing part, 22 ... Sparse matrix storage part, 24 ... Continuous sound spectrogram generation part, 26 ... 1st DESCRIPTION OF SYMBOLS 1 Inverse Fourier-transform part, 30 ... Discontinuous sound processing part, 32 ... Low rank matrix storage part, 34 ... Discontinuous sound spectrogram generation part, 36 ... 2nd inverse Fourier transform part, 38 ... Signal extraction part, 40 ... Robust principal component analysis unit, 41 ... cosine transform unit, 50 ... robust principal component analysis unit, 52 ... inverse cosine transform unit, 53 ... inverse cosine transform unit, 90 ... biological sound signal processing device

Claims (11)

  1. 　生体音信号の原音スペクトログラムを表現した原音行列をロバスト主成分分析でスパース行列と低ランク行列に分解するロバスト主成分分析部と、
    　前記スパース行列を変換して前記生体音信号から連続性の生体音を得る連続性音処理部と、
    　前記低ランク行列を変換して前記生体音信号から連続性の生体音を除外した生体音を得る非連続性音処理部と、
    　を有することを特徴とする生体音信号処理装置。     A robust principal component analysis unit that decomposes an original sound matrix representing an original sound spectrogram of a biological sound signal into a sparse matrix and a low rank matrix by robust principal component analysis;
    A continuous sound processing unit that converts the sparse matrix to obtain continuous biological sound from the biological sound signal;
    A discontinuous sound processing unit that transforms the low rank matrix to obtain a body sound obtained by removing continuous body sounds from the body sound signal;
    A biological sound signal processing apparatus comprising:
  2. 　前記生体音信号を短時間フーリエ変換して前記原音行列を得るフーリエ変換部と、
    　前記非連続性音処理部は、前記低ランク行列から非連続性音スペクトログラムを生成する手段と、前記非連続性音スペクトログラムを短時間逆フーリエ変換して非連続性音信号を生成する手段と、前記非連続性音信号からフーリエ変換信号とウェーブレット信号とを抽出する信号抽出手段とを備えることを特徴とする請求項１に記載の生体音信号処理装置。     A Fourier transform unit that obtains the original sound matrix by performing a Fourier transform on the biological sound signal for a short time;
    The discontinuous sound processing unit generates a discontinuous sound spectrogram from the low rank matrix, a means for generating a discontinuous sound signal by performing a short-time inverse Fourier transform on the discontinuous sound spectrogram, The biological sound signal processing apparatus according to claim 1, further comprising: a signal extraction unit that extracts a Fourier transform signal and a wavelet signal from the discontinuous sound signal.
  3. 　前記生体音信号を短時間フーリエ変換して前記原音行列を得るフーリエ変換部と、
    　前記連続性音処理部は、前記スパース行列から連続性音スペクトログラムを生成する手段と、前記連続性音スペクトログラムを短時間逆フーリエ変換して連続性音信号を生成する手段を備えることを特徴とする請求項１または請求項２に記載の生体音処理装置。     A Fourier transform unit that obtains the original sound matrix by performing a Fourier transform on the biological sound signal for a short time;
    The continuous sound processing unit includes means for generating a continuous sound spectrogram from the sparse matrix, and means for generating a continuous sound signal by performing inverse Fourier transform on the continuous sound spectrogram for a short time. The biological sound processing apparatus according to claim 1 or 2.
  4. 　前記生体音信号を短時間フーリエ変換して前記原音スペクトログラムを得るフーリエ変換部と、
    　離散化した角周波数を行番号とし離散化した時刻を列番号として前記原音スペクトログラムの要素の絶対値を値とする要素からなる原音行列を生成する行列生成部と、
    　をさらに有することを特徴とする請求項１に記載の生体音信号処理装置。     A Fourier transform unit that obtains the original sound spectrogram by performing a Fourier transform on the biological sound signal for a short time;
    A matrix generation unit that generates an original sound matrix including elements whose values are absolute values of elements of the original sound spectrogram, with the discretized angular frequency as a row number and the discretized time as a column number;
    The biological sound signal processing apparatus according to claim 1, further comprising:
  5. 　前記非連続性音処理部は、前記低ランク行列の要素を振幅とし、前記原音スペクトログラムの偏角を偏角とする複素数からなる非連続性音スペクトログラムを生成する手段と、前記非連続性音スペクトログラムを短時間逆フーリエ変換して非連続性音信号を生成する手段と、前記非連続性音信号からフーリエ変換信号とウェーブレット信号とを抽出する信号抽出手段とを備えることを特徴とする請求項４に記載の生体音信号処理装置。 The discontinuous sound processing unit generates a discontinuous sound spectrogram composed of complex numbers having the low rank matrix element as an amplitude and a declination of the original sound spectrogram as an argument, and the discontinuous sound spectrogram 5. A means for generating a discontinuous sound signal by performing inverse Fourier transform on a short time, and a signal extracting means for extracting a Fourier transform signal and a wavelet signal from the discontinuous sound signal. 2. The biological sound signal processing device according to 1.
  6. 　前記連続性音処理部は、前記スパース行列の要素を振幅とし、前記原音スペクトログラムの偏角を偏角とする複素数からなる連続性音スペクトログラムを生成する手段と、前記連続性音スペクトログラムを短時間逆フーリエ変換して連続性音信号を生成する手段を備えることを特徴とする請求項４または請求項５に記載の生体音処理装置。 The continuity sound processing unit generates a continuity sound spectrogram composed of complex numbers having the elements of the sparse matrix as amplitude and the declination of the original sound spectrogram as declination, and the continuity sound spectrogram is inverted for a short time. 6. The biological sound processing apparatus according to claim 4, further comprising means for generating a continuous sound signal by performing a Fourier transform.
  7. 　前記生体音信号を短時間コサイン変換して前記原音行列を得るコサイン変換部と、
    　前記非連続性音処理部は、前記低ランク行列から非連続性音スペクトログラムを生成する手段と、前記非連続性音スペクトログラムを短時間逆コサイン変換して非連続性音信号を生成する手段と、前記非連続性音信号からフーリエ変換信号とウェーブレット信号とを抽出する信号抽出手段とを備えることを特徴とする請求項１に記載の生体音信号処理装置。     A cosine transform unit that obtains the original sound matrix by performing cosine transform on the biological sound signal for a short time;
    The discontinuous sound processing unit includes means for generating a discontinuous sound spectrogram from the low rank matrix, means for generating a discontinuous sound signal by performing a short time inverse cosine transform on the discontinuous sound spectrogram, and The biological sound signal processing apparatus according to claim 1, further comprising: a signal extraction unit that extracts a Fourier transform signal and a wavelet signal from the discontinuous sound signal.
  8. 　前記生体音信号を短時間コサイン変換して前記原音行列を得るコサイン変換部と、
    前記連続性音処理部は、前記スパース行列から連続性音スペクトログラムを生成する手段と、前記連続性音スペクトログラムを短時間逆コサイン変換して連続性音信号を生成する手段を備えることを特徴とする請求項１または請求項７に記載の生体音処理装置。     A cosine transform unit that obtains the original sound matrix by performing cosine transform on the biological sound signal for a short time;
    The continuous sound processing unit includes means for generating a continuous sound spectrogram from the sparse matrix, and means for generating a continuous sound signal by performing inverse cosine transform on the continuous sound spectrogram for a short time. The biological sound processing apparatus according to claim 1 or 7.
  9. 　前記非連続性音処理部は、前記低ランク行列を特異値行列とそれを挟む２つの直交行列との積となるように特異値分解し、前記２つの直交行列から特定の非連続性音の特徴に合致する部分を取り出すことにより前記特定の非連続性音を抽出することを特徴とする請求項１または請求項２、請求項４、請求項７に記載の生体音信号処理装置。 The discontinuous sound processing unit decomposes the low rank matrix into a product of a singular value matrix and two orthogonal matrices sandwiching the low rank matrix, and generates a specific discontinuous sound from the two orthogonal matrices. The biological sound signal processing apparatus according to claim 1, wherein the specific discontinuous sound is extracted by extracting a portion that matches the characteristic.
  10. 　生体音の原音スペクトログラムを表現した原音行列をロバスト主成分分析でスパース行列と低ランク行列に分解するロバスト主成分分析工程と、
    　前記スパース行列を変換して前記生体音信号から連続性の生体音を得る連続性音処理部を得る第２工程と、
    　前記低ランク行列を変換して前記生体音信号から連続性の生体音を除外した生体音を得る第３工程と、
    　を有することを特徴とする生体音信号処理方法。     A robust principal component analysis process for decomposing an original sound matrix representing an original sound spectrogram of a biological sound into a sparse matrix and a low rank matrix by robust principal component analysis;
    A second step of obtaining a continuous sound processing unit for converting the sparse matrix to obtain continuous biological sound from the biological sound signal;
    A third step of converting the low rank matrix to obtain a biological sound excluding continuous biological sounds from the biological sound signal;
    A biological sound signal processing method characterized by comprising:
  11. 　コンピュータを、
    　生体音の原音スペクトログラムを表現した原音行列をロバスト主成分分析でスパース行列と低ランク行列に分解するロバスト主成分分析手段と、
    　前記スパース行列を変換して前記生体音信号から連続性の生体音を得る連続性音処理手段と、
    　前記低ランク行列を変換して前記生体音信号から連続性の生体音を除外した生体音を得る非連続性音処理手段と、
    　として機能させるための生体音信号処理プログラム。     Computer
    A robust principal component analysis means for decomposing an original sound matrix representing an original sound spectrogram of a biological sound into a sparse matrix and a low rank matrix by robust principal component analysis;
    Continuous sound processing means for converting the sparse matrix to obtain continuous biological sound from the biological sound signal;
    Discontinuous sound processing means for converting the low rank matrix to obtain a body sound obtained by excluding continuous body sounds from the body sound signal;
    A biological sound signal processing program for functioning as
PCT/JP2014/078182 2013-10-23 2014-10-23 Biological sound signal processing device, biological sound signal processing method, and biological sound signal processing program WO2015060375A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2015543900A JPWO2015060375A1 (en) 2013-10-23 2014-10-23 Biological sound signal processing apparatus, biological sound signal processing method, and biological sound signal processing program

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2013-220240 2013-10-23
JP2013220240 2013-10-23

Publications (1)

Publication Number Publication Date
WO2015060375A1 true WO2015060375A1 (en) 2015-04-30

Family

ID=52992964

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2014/078182 WO2015060375A1 (en) 2013-10-23 2014-10-23 Biological sound signal processing device, biological sound signal processing method, and biological sound signal processing program

Country Status (2)

Country Link
JP (1) JPWO2015060375A1 (en)
WO (1) WO2015060375A1 (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017204226A1 (en) * 2016-05-23 2017-11-30 国立大学法人京都大学 System and method for target sound signal restoration
CN107704885A (en) * 2017-10-17 2018-02-16 四川长虹电器股份有限公司 Heart sound and the method for lungs sound classification are realized on intelligent platform
US11036981B1 (en) 2020-02-26 2021-06-15 Sas Institute Inc. Data monitoring system
US12009004B2 (en) 2020-02-10 2024-06-11 Tencent Technology (Shenzhen) Company Limited Speech enhancement method and apparatus, electronic device, and computer-readable storage medium

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004357758A (en) * 2003-06-02 2004-12-24 Konica Minolta Medical & Graphic Inc Biological sound detection data processor, biological sound detection data processing method and biological sound detection data processing program

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004357758A (en) * 2003-06-02 2004-12-24 Konica Minolta Medical & Graphic Inc Biological sound detection data processor, biological sound detection data processing method and biological sound detection data processing program

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
TOMONARI MASADA ET AL.: "Feature Extraction by ICA and Clustering for Lung Sound Classification", DBWEB, IPSJ SYMPOSIUM SERIES, vol. 2007, no. 3, 27 November 2007 (2007-11-27), pages 1 - 8 *
TOMOYA SAKAI ET AL.: "Signal Processing for Lung-Sound Recognition: Separation and Classification of Continuous and Discontinuous Adventitious Sounds", IEICE TECHNICAL REPORT, vol. 114, no. 40, 15 May 2014 (2014-05-15), pages 55 - 60 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017204226A1 (en) * 2016-05-23 2017-11-30 国立大学法人京都大学 System and method for target sound signal restoration
JPWO2017204226A1 (en) * 2016-05-23 2019-03-22 国立大学法人京都大学 System and method for recovering acoustic signals
CN107704885A (en) * 2017-10-17 2018-02-16 四川长虹电器股份有限公司 Heart sound and the method for lungs sound classification are realized on intelligent platform
US12009004B2 (en) 2020-02-10 2024-06-11 Tencent Technology (Shenzhen) Company Limited Speech enhancement method and apparatus, electronic device, and computer-readable storage medium
US11036981B1 (en) 2020-02-26 2021-06-15 Sas Institute Inc. Data monitoring system

Also Published As

Publication number Publication date
JPWO2015060375A1 (en) 2017-03-09

Similar Documents

Publication Publication Date Title
US11238881B2 (en) Weight matrix initialization method to improve signal decomposition
Babadi et al. A review of multitaper spectral analysis
Daubechies et al. ConceFT: Concentration of frequency and time via a multitapered synchrosqueezed transform
Li et al. Efficient source separation algorithms for acoustic fall detection using a microsoft kinect
Jin et al. Adventitious sounds identification and extraction using temporal–spectral dominance-based features
WO2015060375A1 (en) Biological sound signal processing device, biological sound signal processing method, and biological sound signal processing program
McDermott et al. Sound texture perception via statistics of the auditory periphery: evidence from sound synthesis
Tsai et al. Blind monaural source separation on heart and lung sounds based on periodic-coded deep autoencoder
Mondal et al. A noise reduction technique based on nonlinear kernel function for heart sound analysis
Emmanouilidou et al. Characterization of noise contaminations in lung sound recordings
Ludeña-Choez et al. Sensor nodes fault detection for agricultural wireless sensor networks based on NMF
Quandt et al. Pulmonary crackle characterization: approaches in the use of discrete wavelet transform regarding border effect, mother-wavelet selection, and subband reduction
De La Torre Cruz et al. Wheezing sound separation based on informed inter-segment non-negative matrix partial co-factorization
Dey et al. Single channel blind source separation based on variational mode decomposition and PCA
Liu et al. Lung sound classification based on Hilbert-Huang transform features and multilayer perceptron network
Ozturk et al. RadioMic: Sound sensing via radio signals
Niegowski et al. Wavelet-based unsupervised learning method for electrocardiogram suppression in surface electromyograms
Dhoulath Beegum et al. Novel implementation of “Filtered-Reconstruction ICA” enabled wheeze component separation for preliminary diagnosis of COVID-19
Acharya et al. Feature extraction techniques for low-power ambulatory wheeze detection wearables
Looney et al. Subspace denoising of EEG artefacts via multivariate EMD
Hedayioglu et al. Separating sources from sequentially acquired mixtures of heart signals
JP6036178B2 (en) Respiratory sound analyzer, respiratory sound analysis method and respiratory sound analysis program
Dubey et al. Does phase matter for monaural source separation?
Andén et al. Classification with joint time-frequency scattering
Wang et al. A step-by-step parameter-adaptive FMD method and its application in fault diagnosis

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 14855176

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2015543900

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 14855176

Country of ref document: EP

Kind code of ref document: A1