JP2005208186A - Apparatus for judging cause - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an apparatus for judging causes which judges causes of abnormality in an objective apparatus or the like according to a plurality of collected data and to accurately judge causes of noise or vibration from a relatively small number of data. <P>SOLUTION: The apparatus is equipped with: a waveform vector generating means which outputs a waveform vector from an input signal; and a judging means which judges the cause of the objective signal to be judged based on the statistic spatial distance between a known waveform vector with a known cause and the objective waveform vector with an unknown cause. The judging means is equipped with: a vector dividing section 39 which divides the known waveform vector and the objective waveform vector into a plurality of small vectors in a desired number; a divided vector distance computing section 40 which computes the plurality of spatial distance groups with respect to the plurality of small vectors; and a comprehensive distance computing section 41 which computes the comprehensive spatial distance in a cause-classified reference space by regarding the sequence of the plurality of spatial distance groups as new vectors. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a cause determination device for determining the cause of an abnormality of a target device or the like based on a plurality of collected data, in particular, sound emitted from a machine, particularly impact sound generated from office equipment such as a copying machine or a printer. The present invention relates to a cause determination apparatus suitable for identifying the cause of occurrence of a problem.

  2. Description of the Related Art Conventionally, there is known a method for estimating and determining the cause of occurrence based on signals due to sound or vibration. On the other hand, for example, there is a need to improve the office environment in order to support improvement in comfort or intellectual productivity in the office, and office equipment such as copiers and printers is required to reduce the noise. To that end, it is important to first identify the cause of noise.

  FIG. 9 shows a time waveform of noise generated during operation of a general electrophotographic printer. Here, a portion indicated by a circle is a portion in which a sound called an impact sound that shows a high fluctuation in sound pressure level is generated in a short time. Such a sound is generated due to a collision between components, a collision between paper and a component, and is generated, for example, in a solenoid or a paper positioning unit during paper running.

In general, the impact sound shows high sound pressure fluctuation, so that it is easy for humans to recognize and has a startle effect. Therefore, the developer of the apparatus performs an activity for reducing the impact sound, but it is very difficult to determine the cause of the impact sound, particularly in office equipment such as a copying machine and a printer. This is because mechanical noise such as copying machines and printers has a feature that various sound sources are operated at various timings. For the method of identifying the sound source of stationary sound or impact sound generated from the device,
(A) a method of measuring by sound intensity, etc.
(B) Check the operation timing of parts that may be sound sources,
(C) Check if there is no impact sound by turning off the operation of parts that may be sound sources,
(D) Check if the impact sound is reduced by attaching a material that absorbs shock to a part that may be a sound source or reducing the collision speed.
(E) Identify the sound source by examining the feature quantity by frequency spectrum analysis, etc.
There exist methods such as.

  However, copying machines, printers, and the like are miniaturized, and each component is stored in a very close position. For this reason, (a) the spatial resolution is insufficient to specify a component by a measurement method such as acoustic intensity. In addition, copying machines, printers, and the like have been speeded up, and each component is driven almost simultaneously. For this reason, (b) the time resolution is insufficient to identify the component by the method of checking the operation timing of the component. Therefore, at present, (c) a method for confirming whether or not the collision sound is lowered by turning off the operation of a component that may be a sound source, or (d) pasting a material that absorbs an impact to a component that may be a sound source. A method of confirming whether the impact sound is reduced by lowering the collision speed is adopted.

  In such a method for identifying the cause of the impact sound ((c), (d)), it takes a lot of time and labor to reduce the impact sound, which is the final target. Impact sounds may remain. In the future, as the device becomes smaller and more advanced, the number of parts that can be a sound source becomes smaller and more diverse, and a method for identifying the cause of the occurrence of such an impact sound ((c), (d )), It becomes even more difficult to deal with it.

  In order to solve such a problem, there has been proposed a method for estimating the cause of sound generation by analyzing the qualitative characteristics of the sound rather than clarifying the position where the sound is generated. For example, (e) frequency spectrum analysis or the like is effective for analysis of stationary sound. On the other hand, detailed frequency information cannot be obtained for an impact sound generated from a copying machine, a printer, or the like because analysis time is short. As an attempt to estimate the cause of an occurrence by applying it to an impact sound, for example, Patent Document 1 discloses a sound source type identification device that determines the type of a sound source using power spectrum analysis and a neural network. Such an analysis focusing on sound quality is considered to be a promising approach that can solve the problem of insufficient resolution due to the dependence on the wavelength of the sound and the distance to the sound source.

  However, this sound source type identification device also has the following drawbacks. According to this sound source type identification device, it is possible to distinguish sound sources that are characteristic in frequency.For example, when the support form of the metal plate changes with respect to the impact sound that occurs when metal plates of the same material collide with each other, etc. It is known that characteristics on the time axis such as reverberation time change greatly without changing the frequency characteristics of the generated sound. Can not. Further, since the time from the generation of the impact sound to the disappearance is as short as several tens to several hundreds of milliseconds, there is a problem that it is basically difficult to apply frequency analysis such as FFT analysis.

  In order to solve this problem, Patent Document 2 discloses a sound source recognition apparatus that employs measurement of a difference in duration of sound sources in addition to identification of frequency components of sound sources. In this sound source recognition device, it is possible to identify the difference due to the difference in the support form as described above by measuring the difference in the duration, but the temporal characteristic of the impact sound is only the duration. Rather, for example, it has been reported that if the surface hardness of the impact member is different even at the same duration, a large difference occurs in the transient characteristics such as the rising curve and attenuation curve of the sound pressure amplitude. It is not possible to discriminate accurately by comprehensively judging various features appearing on the time characteristics only by measuring the.

  Similarly, Patent Document 3 calculates and extracts a time waveform and a frequency waveform of a sound to be evaluated to generate a waveform vector, and calculates a spatial distance with respect to a feature vector for each cause obtained in advance as a known waveform vector. Discloses a signal determination device for determining the cause of occurrence. According to this signal determination device, it is possible to comprehensively determine temporal and frequency characteristics of a sound source using a statistical measure called spatial distance, and highly accurate sound source analysis is possible.

However, it has become clear that this signal determination apparatus has the following drawbacks. This is due to the use of a statistical index called spatial distance, and the number of data samples (number of cases) constituting the feature vector is sufficient compared to the number of elements constituting the waveform vector, that is, the number of waveform points. If it is not large, the degree of freedom in statistical processing is insufficient, and a correct spatial distance value cannot be calculated. Specifically, it is necessary to secure the number of cases more than three times the number of waveform points. For example, if the number of waveform points is 30, the number of cause-specific data samples for constructing the feature vector may be 90 or more, but if the number of data points is 500, 1500 or more data samples are required. If a large number of data samples cannot be collected, the number of waveform points must be reduced. In this case, even if you want to perform signal analysis using both time waveform and frequency waveform, there is a restriction such as giving up the adoption of frequency waveform data due to the number of waveform points, and sufficient analysis accuracy is achieved. It will cause problems such as inability to obtain.
JP 2000-275096 A JP-A-9-81180 JP 2002-323370 A

  The present invention has been made in view of the above technical problems, and an object of the present invention is to provide a cause determination apparatus that can accurately determine the cause of occurrence of sound, vibration, and the like with a relatively small number of data. It is to provide.

The cause determination apparatus of the present invention that achieves the above object is
A vector generation unit that generates a vector including a plurality of data generated by generating a plurality of data based on the collected plurality of data; and
Based on a plurality of data collected a plurality of times at a stage where the cause is known, a reference space for each cause is obtained based on a plurality of known vectors generated by the vector generation unit, and a plurality of data collected in a state where the cause is unknown A determination unit that obtains a statistical distance in the reference space of the determination target vector generated by the vector generation unit based on the data of the reference and determines a cause corresponding to the determination target vector based on the statistical distance In the cause determination device comprising
The above judgment part
A vector dividing unit that divides the vector generated by the vector generating unit into a plurality of small vectors each of which includes elements when the elements of the vector are divided into a plurality of groups;
Based on a plurality of known small vectors obtained by dividing a plurality of known vectors by the vector dividing unit, 1 for each of a plurality of known small vectors whose elements are mutually corresponding groups constituting a plurality of known vectors of the same cause Two first reference spaces, and for each of a plurality of determination target small vectors obtained by dividing the determination target vector by the vector dividing unit, each statistic on each first reference space corresponding to each determination target small vector A divided vector distance calculation unit that obtains a determination target distance vector having a plurality of statistical distances corresponding to the plurality of determination target small vectors as elements,
A second reference space is obtained based on a plurality of known distance vectors having a statistical distance on each first reference space corresponding to each of the plurality of known small vectors as an element; An overall distance calculation unit for obtaining a statistical distance on the second reference space of the determination target distance vector obtained by the divided vector distance calculation unit;
And a final determination unit that determines a cause corresponding to the determination target vector based on a statistical distance in the second reference space obtained by the total distance calculation unit.

  Here, the vector generation unit may generate a vector based on a plurality of data arranged in time series obtained by collecting operating sounds of the target device, in which case the determination unit The cause of the noise of the target device may be determined.

  An image forming apparatus that transports a recording medium and forms an image on the recording medium that is being transported can be the target apparatus.

  Furthermore, the vector generation unit includes a signal extraction unit that detects predetermined reference data from a plurality of data arranged in time series and extracts a plurality of data including data temporally preceding the reference data. It is preferable.

  Further, the vector generation unit may generate a time waveform vector having a plurality of data representing a time waveform based on a plurality of data arranged in time series, or the vector generation unit may A plurality of data arranged in time series may be converted into a plurality of data in the frequency domain to generate a frequency waveform vector having a plurality of data representing a frequency distribution as elements, or the vector generation unit Is a plurality of data representing a time waveform obtained based on a plurality of data arranged in time series, and a plurality of data in the frequency domain obtained by converting the plurality of data arranged in the time series into a plurality of data in the frequency domain. It is also possible to generate a comprehensive waveform vector having both the data and the data as elements.

  Furthermore, it is preferable that the division distance calculation unit and the total distance calculation unit calculate a Mahalanobis distance as a statistical distance.

  When causing the cause determination apparatus of the present invention to determine the cause, a vibration signal or an acoustic signal may be input, and a plurality of data representing the vibration or sound may be collected. Here, in the case of an acoustic signal, the vector generation unit may include an auditory correction unit that corrects an auditory sense of the input acoustic signal using a predetermined correction filter. As specific examples of correction filters, A, B, C, D, and E characteristic filters can be used, such as a high-pass filter (to remove low-frequency background noise components) and other unique filters. A characteristic filter can also be used. In general, it is preferable to use an A characteristic filter that is close to human hearing.

  Moreover, what is suitable for the determination by the abnormality determination device according to the present invention is an acoustic signal having a duration of 1000 [msec] or less, further 100 [msec] or less, such as a collision sound (between objects) and an impact sound. It is. There are no particular restrictions on the signal source, but in view of the fact that there are countless moving parts in the device and the fact that it is often installed in a place where a relatively quiet environment is required, the signal source However, when an image forming apparatus (such as a copying machine, a printer, or a facsimile) is used as a target apparatus, it is preferable to apply the present invention because a great effect can be expected for the noise reduction of the image forming apparatus.

  Examples of the signal from the image forming apparatus include a collision acoustic signal between components inside the image forming apparatus and an acoustic signal caused by a recording medium inside the image forming apparatus. More specifically, the collision acoustic signal between parts includes (a) collision sound between metal parts, (b) collision sound between resin parts, (c) collision sound between metal parts and resin parts, and the like. As acoustic signals caused by the recording medium, (d) a collision sound between the recording medium and a component, (e) a buckling sound of the recording medium, (f) a sound in which the rear end in the conveyance direction of the recording medium is bounced, etc. Can be mentioned.

  The vector generation unit may include a signal extraction unit that outputs a waveform vector from a part of the input signal. In this case, the signal cutout unit can cut out a part of the input signal based on the rising part, maximum part or maximum part of the input signal, and cut out the signal after the reference in terms of time. It is possible to cut out the signal retroactively from the reference.

  Further, the vector generation unit may include an amplitude standardization unit that standardizes the maximum value or the minimum value of the collected data.

  The vector generation unit may include a time waveform vector generation unit that outputs a time waveform vector in the time domain of the input signal, and a frequency waveform vector generation unit that outputs a frequency waveform vector in the frequency domain of the input signal. May be provided. Further, the vector generation unit includes a time waveform vector generation unit that outputs a time waveform vector in the time domain of the input signal, a frequency waveform vector generation unit that outputs a frequency waveform vector in the frequency domain of the input signal, A total waveform vector generation unit that combines the time waveform vector and the frequency waveform vector and outputs a total waveform vector may be provided.

  Here, examples of the time waveform vector in the time domain include those having the sound pressure level for each predetermined time as each vector component. In order to obtain a time waveform vector whose sound pressure level at each predetermined time is a vector component, a time waveform vector generation unit is configured by software, or a sound pressure leveling circuit used in a general sound level meter is used for time. A waveform vector can also be obtained. The predetermined time can generally be set at regular intervals, but can also be set at arbitrary intervals. In addition, examples of the frequency waveform vector in the frequency domain include those having values for each predetermined frequency obtained by frequency analysis as vector components.

  For example, the vector generation unit is input with a time waveform vector generation unit that outputs a time waveform vector having a sound pressure value for each predetermined time as each vector component as a time waveform vector in the time domain of a plurality of collected data. A frequency waveform vector generation unit that outputs a frequency waveform vector having a power spectrum value for each predetermined frequency subband as each vector component as a frequency waveform vector in the frequency domain of the signal to be transmitted, the time waveform vector, and the frequency waveform vector In addition, a total waveform vector generation unit that outputs a total waveform vector may be provided.

  In addition, although every predetermined frequency generally means every predetermined frequency range (subband), the frequency range may be a fixed range or may be an arbitrary range. Specific examples of frequency analysis include FFT (Fast Fourier Transform) analysis, wavelet analysis, and generalized harmonic analysis.

  The frequency waveform vector generation unit may include a window function processing unit that performs window function processing on a time waveform that is a target of frequency analysis. Examples of the window function include a rectangular window, a Hanning window, a Hamming window, a Blackman window, and a flat top window.

  The determining unit includes a vector dividing unit that divides the known waveform vector and the determination target waveform vector into an arbitrary plurality of small vectors, a divided vector distance calculating unit that calculates a plurality of spatial distance groups for each of the plurality of small vectors, A small vector component in which each known waveform vector is divided into a plurality of components, each of which includes a total distance calculation unit that calculates a total spatial distance in the second reference space for each cause by using a sequence of spatial distance values as a new vector. And the first spatial distance group is calculated with each of the small vector components obtained by dividing the judgment target waveform vector into a plurality of variables as variables, and the first spatial distance group is used as a new vector component to determine the known waveform vector. A second spatial distance value representing the relationship with the waveform vector is calculated to determine the relationship between the two. As a statistical spatial distance measure, a general distance measure used in discriminant analysis or cluster analysis can be used. Specific examples of the distance measure include Euclidean distance, standardized Euclidean distance, Minkowski distance, Mahalanobis distance, and the like.

  As described above, according to the present invention, the determination unit includes the vector dividing unit, and the statistical distance is calculated in two stages by including the divided vector distance calculating unit and the total distance calculating unit when calculating the statistical distance. Therefore, the number of data for determining the cause is greatly reduced as compared with the conventional case, and more accurate determination is performed.

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

  FIG. 1 is a diagram showing a cause determination system including a cause determination apparatus according to an embodiment of the present invention.

  The cause determination system 100 shown in FIG. 1 uses a copying machine (image forming apparatus) T as a determination target, collects noise (acoustic signal) emitted from the copying machine, and causes the abnormal sound generated by the copying machine T. Is to judge.

Here, the cause of the noise emitted from the copier T is
Cause (1): Impact noise between metal parts,
Cause (2): Impact sound between resin parts,
Cause (3): Impact sound between metal parts and resin parts,
Cause (4): Impact sound between recording medium and components,
Cause (5): Buckling sound of recording medium absence,
Cause (6): A sound that the rear end of the recording medium in the transport direction is bounced,
Is determined.

  The entire cause determination system 100 includes a macrophone 1 as sound collection means for capturing noise from the copier T, and a DAT (digital audio tape) recorder 2 as analog / digital conversion means connected to the microphone 1. And a personal computer system C connected to the DAT recorder 2. The personal computer system C further includes a computer main body 3, a keyboard 4a as an input means, a mouse 4b, a display device 5 as an output means, and the like.

  The hardware resources in the computer main body 3 include a CPU as arithmetic control means, a RAM as main storage means, a hard disk as auxiliary storage means, an input / output control device (not shown), and the like. Software resources in the main body 3 include an operating system, acoustic analysis software, numerical analysis software, and the like (all not shown). The functions of the vector generation unit G and the determination unit 3J shown in FIG. 2 are realized by the joint work of these hardware resources and software resources.

  FIG. 2 is a basic functional block diagram of the cause determination system 100. The functions of the cause determination system 100 include a microphone 1, a DAT recorder 2, a vector generation unit 3G, a determination unit 3J, and a display unit 5 in order along the processing flow. The signals input / output between these basic function blocks are an analog signal AS, a digital signal DS, a waveform vector V, and a determination result R.

  3 is a detailed functional block diagram showing the configuration of the vector generation unit 3G shown in FIG. 2, and FIG. 4 is a detailed functional block diagram showing the configuration of the determination unit 3J shown in FIG.

  As shown in FIG. 3, the vector generation unit 3G includes a data storage unit 30, a time waveform vector generation unit 3T, a frequency waveform vector generation unit 3F, and an overall waveform vector generation unit 36. The time waveform vector generation unit 3T includes an auditory correction unit 31, a cutout unit 32, and a standardization unit 33. The frequency waveform vector generation unit 3F includes a window function in addition to these units 31, 32, and 33. A processing unit 34 and a frequency analysis unit 35 are provided. The signals input / output between these detailed functional blocks are a digital signal DS, a hearing-corrected digital signal DS ′, a cut-out digital signal ΔDS ′, and a standardized digital signal “ΔDS ′” (= time waveform). Vector TV), window function processed digital signal “ΔDS ″”, frequency analyzed digital signal (= frequency waveform vector FV), and total waveform vector V.

As shown in FIG. 4, the determination unit 3J includes a Mahalanobis distance calculation units 37 (a) to (f) corresponding to a plurality of cause groups, here, six types (a) to (f), and a final determination unit 38. And. The signals input / output between these detailed functional blocks are the comprehensive waveform vector V, the Mahalanobis distances D2 (a) to (f) from the respective cause vector spaces, and the determination result R. The Mahalanobis distance calculation unit 37 for each cause includes a vector dividing unit 39 that divides the total waveform vector V into a plurality of (here, z) small vectors, and a divided vector distance that calculates the spatial distance of the divided small vectors. Calculation units 40 (1) to 40 (z), and a total distance calculation unit 41 for calculating a total spatial distance from a group of spatial distances for each division vector, and divided vectors V (1) to V (z), The divided vector distances D ² (1) to D ² (z) corresponding to this, and the Mahalanobis distance D2 obtained by combining them are calculated.

For example, in the cause space (a), the total waveform vector V is divided into small vectors V (a1) to V (az) by the vector dividing unit 39 (a), and divided vector distance calculating units 40 (a1) to 40 (40). In (az), the divided vector distances D ² (a1) to D ² (az) corresponding to the respective small vectors V (a1) to V (az) are calculated, and these are integrated by the total distance calculation unit 41 (a). The Mahalanobis distance D ² (a) is calculated. The Mahalanobis distances D ² (b) to D ² (f) are similarly calculated in the other cause spaces (b) to (f).

  5 and 6 are flowcharts for explaining the basic usage method and operation of the cause determination system 100. FIG. Hereinafter, based on this flowchart, the operation of the cause determination system 100 according to the present embodiment will be described.

  First, prior to the cause determination, pre-processing is performed in advance for the cause determination system 100 (step S1 in FIG. 5). In this preprocessing, a plurality (m) of samples y of the total waveform vector V are obtained for each cause (a) to (f), and an inverse matrix A is obtained for each cause (a) to (f). The data is stored in the Mahalanobis separation calculation units 37 (a) to (f) (see FIG. 4).

In the present embodiment, m = 200 because 200 sampled sound signals are captured for one cause. In the present embodiment,
Cause (a): Impact sound between metal parts,
Cause (b): Impact sound between resin parts,
Cause (c): Impact sound between metal part and resin part,
Cause (d): Impact sound between recording medium and component parts,
Cause (e): Buckling sound of the recording medium,
Cause (f): A sound that the rear end in the transport direction of the recording medium is bounced,
6 types of causes are judged. Further, as will be described later, the dimension k of each comprehensive waveform vector is k = 256 + 128 = 384.

  A method for obtaining the inverse matrices A and B will be described later.

  When the preprocessing (step S1 in FIG. 5) ends, the cause of occurrence is determined based on the acoustic signal whose cause is unknown (steps S2 to S6 in FIG. 6).

  First, the microphone 1 measures an acoustic signal to be determined (the cause of occurrence is unknown) (step S2 in FIG. 6). In this case, noise from the copying machine T is measured by the microphone 1 installed in the vicinity of the copying machine T.

  The noise from the copying machine T is a sound wave, which vibrates the vibrator in the microphone 1 and converts the vibration into an analog signal AS (see FIG. 2).

Next, the analog signal AS is digitized by the DAT recorder 2 (step S3 in FIG. 6). This is performed by converting the analog signal AS into the digital signal DS by the A / D conversion circuit in the DAT recorder 2 connected to the microphone 1 (see FIG. 2). Further, the digital signal DS is once recorded in the DAT recorder 2. FIG. 9 is a graph showing the digital signal DS recorded in the DAT recorder 2. The vertical axis of this graph represents sound pressure [× 10 ⁻⁶ Pa], and the horizontal axis represents time [sec].

  In this embodiment, since the resolution (sampling frequency) of the A / D conversion circuit of the DAT recorder 2 is 48 [kHz], the obtained digital signal DS is obtained as 48000 points of sampling data per second. Here, it is desirable to use a sampling frequency of at least 10 [kHz] or more, preferably 20 [kHz]. This is because the characteristic of impact sound generated from an image forming apparatus such as the copying machine T or printer often appears in a high frequency band of 5 to 10 kHz or more. In order to analyze a band of 5 to 10 [kHz] or more, the sampling frequency needs to be 10 to 20 [kHz] or more, which is twice that of the sampling theorem.

  Next, the vector generation unit 3G generates a waveform vector V corresponding to the acoustic signal to be determined (step S4 in FIG. 6). First, the data storage unit 30 stores the digital signal DS recorded and output by the DAT recorder 2. Specifically, it is stored in the hard disk of the personal computer main body 3 connected to the DAT recorder 2. In the present embodiment, the digital signal DS including the impact sound (the portion where the sound pressure rapidly increases) indicated by the circle in FIG. 9 is stored for about 0.01 seconds, that is, about 480 sampling data.

  FIG. 10 is a graph showing the digital signal DS stored in the data storage unit 30.

  The vertical axis of this graph represents the sound pressure, and the horizontal axis represents the number of sampling data.

  Next, the auditory correction unit 31 performs auditory correction on the digital signal DS stored in the data correction unit 30. This corrects the digital signal DS stored by the correction filter preset by the auditory sensation correction unit 31 to obtain a new digital signal DS ′ (see 3). In the present embodiment, correction is performed using an A filter that is close to human hearing.

  Next, the cutout unit 32 cuts out a part of the audible corrected digital signal DS ′. This cuts out a part of the auditory corrected digital signal DS ′ based on the conditions set in the cutout unit 32 to obtain a cutout digital signal ΔDS ′ (see FIG. 3). In the present embodiment, the condition set in the cutout unit 32 is based on the rising part of the audible corrected digital signal DS ′ as a reference, and from the data traced back in time by 50 sampling data with respect to the reference. Up to 205 sampling data with respect to the reference are cut out in time (that is, 50 + 1 + 205 = 256 data). As a method for detecting the rising portion, the maximum value of the absolute value of the sound pressure p of the digital signal DS ′ is detected, and the portion that reaches the sound pressure value of several percent of the maximum value for the first time is defined as the rising portion. can do. In this embodiment, the cutout unit 32 is realized by software. However, for example, a part of the digital signal DS ′ whose auditory sense has been corrected can be cut out by a data logger having a trigger function.

  FIG. 11 is a graph showing the digital signal ΔDS ′ cut out by the cutout unit 32. The vertical axis of this graph represents the sound pressure, and the horizontal axis represents the number of sampling data.

  In FIG. 11, p (MAX) represents the absolute maximum value of the sound pressure p, and p (TH) represents the sound pressure value at the rising portion.

  From the cut-out digital signal ΔDS ′ obtained in this way, the original time waveform vector Tv can be considered. That is, the time waveform original vector Tv = (p1, p2,..., P51 (= p (TH)),..., P () having the sound pressure values p1 to p256 of the 1st to 256th sampling data as elements of the vector. MAX),..., P256) can be obtained.

  Next, the standardization unit 33 standardizes the value of the cut-out digital signal ΔDS ′. This is based on the standard value preset in the standardization unit 33 and the maximum value of the absolute value of the sound pressure included in the digital signal ΔDS ′, and standardizes the value of the digital signal ΔDS ′. A digital signal “ΔDS ′”, that is, a time waveform vector TV is obtained (see FIG. 3). This time waveform vector TV is based on the previous time waveform original vector Tv, the maximum sound pressure value p (MAX), and the reference value S. TV = (S / p (MAX)) × Tv = (S / p (MAX)) × (p1, p2,..., P51 (= p (TH)),..., P (MAX),..., P256) = ((S / p (MAX)) × p1, (S / p (MAX)) × , (S / p (MAX)) × p51,..., (S / p (MAX)) × p (MAX)) (= S),..., (S / p (MAX)) × p256) = (P1,..., P256).

  Next, the window function processing unit 34 performs window function processing on the standardized digital signal “ΔDS ′” (= time waveform vector TV). This is to process the standardized digital signal “ΔDS ′” by the window function set in advance in the window function processing unit 34 to obtain the window function processed digital signal “ΔDS ″” (FIG. 3). reference). In this embodiment, the Hanning window function is used as the window function. By performing such processing, it is possible to reduce the influence of a folding error or a leakage error.

  Next, the frequency analysis unit 35 performs frequency analysis on the window function processed digital signal “ΔDS ″”. In this case, the frequency analysis unit 35 performs FFT analysis on the window function-processed digital signal “ΔDS ″” to obtain a frequency spectrum for each frequency subband.

FIG. 12 is a graph showing the FFT processing result by the frequency analysis unit 35. In this graph, the vertical axis represents the power spectrum [× 10 ⁻⁶ Pa], and the horizontal axis represents the frequency (subband). In the present embodiment, FFT analysis is performed based on the 256 digital signals “ΔDS ″” described above, and 128 new digital signals indicated by black dots in FIG. 12 are obtained.

  The frequency waveform vector FV can be considered from the digital signal obtained in this way (see FIG. 3). That is, the frequency vector FV = (ps1, ps2,..., Ps128) having the first to 128th power spectrum ps values in order from the lowest frequency as each element of the vector is obtained.

  Next, the comprehensive waveform vector generation unit 36 obtains a waveform vector V. Here, a 256-dimensional time waveform vector TV = (P1,..., P256) generated by the above-described time waveform generation unit 3T, and a 128-dimensional frequency waveform vector FV generated by the above-described frequency waveform generation unit 3F = From (ps1,..., Ps128), 256 + 128 = 384-dimensional waveform vector V = (P1,..., P256, ps1,..., Ps128) = (v1,..., V384) is obtained (see FIG. 3). It should be noted that the waveform vector V = (v1,..., V384) obtained in this way is obtained from a sample acoustic signal whose cause is clear in advance, and the known waveform vector y = (v1,..., V384) = (y1,. .., Yk), and what is obtained from an acoustic signal to be determined with an unknown cause is referred to as a determination target waveform vector x = (v1,..., V384) = (x1,..., Xk) (k = 384). .

  Next, the determination unit 3J determines the cause of generation of the acoustic signal to be determined using the Mahalanobis distance, and obtains a cause determination result R (step S5 in FIG. 6).

  Here, it returns to description of step S1 of FIG.

The Mahalanobis distance D ² of the evaluation target waveform vector (x ₁ ,..., Z _k ) in a certain known waveform vector space is the inverse matrix of the correlation coefficient matrix R relating to the data group (number of cases m) of the known waveform vector (k term). Is calculated from the following equation (1). Here, the correlation coefficient r constituting the correlation coefficient matrix R is calculated as follows. First, the known waveform vector y = (y1,..., Yk) × m sets (cases) of data is rearranged into m case data groups for each of 1 to k components constituting the known waveform vector, Y1 = (y11 to y1m), Y2 = (y21 to r2m),..., Yk = (yk1 to ykm) are generated. The correlation coefficient between each data group is calculated, the correlation coefficient between Y1 and Y2 is r12, the correlation coefficient between Y1 and Y3 is r13, the correlation coefficient between Y1 and Yk is r1k, and the correlation coefficient between Y2 and Yk Is a correlation coefficient matrix R.

但し、

However,

  In order to obtain the inverse matrix A of the correlation coefficient, it is necessary to ensure k <m, that is, the number of cases larger than the number of waveform vector terms. Furthermore, in order to calculate the Mahalanobis distance with sufficient accuracy, it is desirable that 3 × k <m. FIG. 7 shows the result of calculating the Mahalanobis distance by actually changing the number of terms and the number of cases in the waveform vector. Here, for the same cause group data, the number of cases was fixed at 3200 and the number of waveform vector terms was changed. Specifically, the Mahalanobis distance is calculated and the average value is obtained using the 320 case data groups as the known waveform vector group and each of the 320 case data as the evaluation target waveform vectors. Since the known waveform vector group and the waveform vector to be evaluated are the same, these spatial distances are supposed to be minimized and should be as close as possible to the center of gravity position “1”, but the number of terms in the vector is the number of cases. It can be seen that the spatial distance increases rapidly when the distance is close to. In this state, it is determined that these data are completely different, and the determination is wrong. This is because statistical redundancy in calculating the inverse matrix A is lost. From FIG. 7, with respect to the number of cases 320, if the number of vector terms is 75-100 or less, the spatial distance can be regarded as almost “1”.

  That is, when the number of cases more than three times the number of terms in the waveform vector cannot be secured, it cannot be determined with sufficient accuracy. Therefore, in this embodiment, this problem is avoided by dividing the waveform vector into a size that satisfies 3 × k <m. If the number of divisions is z, in this embodiment, k = 384 and m = 200. Therefore, in order to satisfy 3 × k <m, the number of divisions z is desirably 6 or more. Here, assuming that z = 6, the vector dividing unit 39 divides the known waveform vector into six small vectors. For example, for the cause space (a), the vector dividing unit 39 (a) divides the known waveform vector ya = (v1,..., V384), and v (a1) = (v1,..., V64), v (a2). ) = (V65,..., V128), v (a3) = (v129,..., V192), v (a4) = (v193,..., V256), v (a5) = (v257,..., V320), v Six small vectors of (a6) = (v321,..., V384) are generated. Each small vector has 64 terms and 200 cases, which satisfies the condition of 3 × k <m. In step S1 of FIG. 5, the inverse matrix of the correlation coefficient matrix is obtained for each of v (a1) to v (a6) using Equation (1), and the first inverse matrix A (a1) to A (a6) is stored in the divided vector distance calculation units 40 (1) to (z) of the Mahalanobis distance calculation unit 37.

In step S1 in FIG. 5, the Mahalanobis distance is calculated by regarding each of the known waveform vectors (number of terms k = 384) as the judgment target waveform vectors for all cases m = 200. Specifically, in the same manner as described above, the waveform vector ya = (v1,..., V384) is divided by the vector dividing unit 39 (a), and v (a1) = (v1,..., V64), v (a2). = (V65, ..., v128), v (a3) = (v129, ..., v192), v (a4) = (v193, ..., v256), v (a5) = (v257, ..., v320), v ( 6 small vectors of a6) = (v321,..., v384) are generated. With respect to v (a1) to v (a6), the divided vector distance calculation units 40 (a1) to 40 (a6) use the first inverse matrices A (a1) to A (a6), and the Mahalanobis distance D ^2. (A1) to D ² (a6) are obtained. This calculation is repeated for the number of cases m = 200, and 200 sets of Mahalanobis distances D ² (a1) to D ² (a6) are generated. Next, the data group of D ² (a1) to D ² (a6) × 200 sets is regarded as a second known vector group of k = 6 and m = 200, and the same calculation is performed by the same operation as the inverse matrix A described above. An inverse matrix of the relational number matrix is obtained and stored as the second inverse matrix B (a) in the total distance calculation unit 41 (a) of the Mahalanobis distance calculation unit 37 (a). From k = 6 and m = 200, the condition of 3 × k <m is also satisfied here. This is also performed for the cause spaces (b) to (f), and finally the first inverse matrices A (a1) to A (a6), ..., A (f1) to A (f6), and the second inverse The matrices B (a),..., B (f) are stored in the Mahalanobis distance calculation unit 37, and the preprocessing ends.

  Next, the Mahalanobis distance calculation procedure of the judgment target waveform vector will be described (step S5 in FIG. 6).

First, the judgment target waveform vector x = (v1,..., X384) = (x1,..., X386) is divided by the vector dividing unit 39 (a), and x (1) = v (a1) = (x1,. ), X (2) = v (a2) = (x65,..., X128), x (3) = v (a3) = (x129,..., X192), x (4) = v (a4) = (x193) ,..., X256), x (5) = v (a5) = (x257,..., X320), x (6) = v (a6) = (x321,..., X384) are generated. The With respect to v (a1) to v (a6), the divided vector distance calculation units 40 (a1) to 40 (a6) use the first inverse matrices A (a1) to A (a6), and the Mahalanobis distance D ^2. (A1) to D ² (a6) are obtained. Next, using the second inverse matrix B (a) stored in advance in the Mahalanobis distance calculation unit 37 (a), the total distance calculation unit 41 (a) uses the Mahalanobis distances D ² (a1) to ^D 2 ( The final Mahalanobis distance D ² (a) is calculated by regarding the sequence of a6) as a vector.

Hereinafter, similarly, the same calculation is repeated for the cause spaces (b) to (f), and the Mahalanobis distances D ² (b) to D for each cause in the Mahalanobis distance calculation units 37 (b) to (f). ² (f) is calculated (see FIG. 4).

Next, the final determination unit 38 selects the one having the smallest Mahalanobis distance D ² (a) to D ² (f), and sets the cause corresponding to the Mahalanobis distance as the determination result R. For example, when D ² (a) is the smallest among Mahalanobis distances D ² (a) to D ² (f), the acoustic signal to be determined is “cause (1): collision sound between metal parts” "

  Next, the determination result R is displayed on the display unit 5 (step S6 in FIG. 6). For example, if the acoustic signal to be determined is “Cause (1): Collision sound between metal parts”, that fact is displayed on the screen of the display unit 5 (see FIG. 2).

  The present inventors have conducted a number of experiments to complete the present invention. Here, a part of the experiments will be described as examples to confirm the effect of the present invention. In this example, the cause determination system 100 according to the above-described embodiment is used, and cause (2): collision sound between resin (plastic) parts and cause (3): between metal parts and resin (plastic) parts. The judgment of collision sound was verified. The cause (2) and the cause (3) are generally similar in audible impression and sound pressure waveform, and are relatively difficult to distinguish.

  First, 100 waveform vectors V (3) (number of terms k = 384) (v1 (3) to v100 (3)) are obtained from the collision sound due to the cause (3) (steps S2 to S4 in FIG. 6), and these Of the 100 waveform vectors V (3), 80 waveform vectors V (3) selected at random are known waveform vectors y (3) (y1 (3) to y80 (3)) (m = 80). ), And the remaining 20 judgment object waveform vectors x (x1 to x20). Next, 100 waveform vectors V (2) (v1 (2) to v100 (2)) are obtained from the collision sound due to the cause (2) (steps S2 to S4 in FIG. 6), and all of these (for verification) are obtained. ) It is assumed that the determination target waveform vector x (x21 to x120).

Next, the known waveform vectors y1 (3) to y80 (3) are divided into 12 small vectors (z = 12) so as to satisfy 3 × k <m, and the first inverse matrix A (3, 1) ~ A (3,12) and the second inverse matrix B (3) are calculated, and the Mahalanobis distance calculation unit 37 (3) stores them (see FIG. 4). Next, the first inverse matrices A (3, 1) to A (3) obtained by the preprocessing by the Mahalanobis distance calculation unit 37 (3) for the determination target vectors x1 to x20 and x21 to x120 (for verification) , 12), and the respective Mahalanobis distances D ² 1 (3) to D ² 20 (3) and D ² 21 (3) to D ² 120 (3) based on the second inverse matrix B (3) To do.

FIG. 8 is a graph showing the results of this experiment. In this graph, the vertical axis represents the Mahalanobis distance D2 (3), and the horizontal axis represents the number of waveform vectors (number of samples). Here, the Mahalanobis distances D ² 1 (3) to D ² 20 (3) finally obtained from the collision sound due to the cause (3) are all close to 1 (all are 2 or less). On the other hand, the Mahalanobis distances D ² 21 (3) to D ² 120 (3) finally obtained from the collision sound due to the cause (2) all have large values (approximately several hundred to about 1000). .

Therefore, according to the present embodiment, the cause (2): the collision sound between the resin (plastic) parts and the cause (3): the collision sound between the metal part and the resin (plastic) part are generally difficult to distinguish. Can be clearly distinguished by setting the threshold of the Mahalanobis distance D ² (3) to 10, for example.

It is a figure which shows the cause determination system containing the cause determination apparatus of one Embodiment of this invention. It is a basic functional block diagram of a cause determination system. FIG. 3 is a detailed functional block diagram illustrating the vector generation unit illustrated in FIG. 2 in more detail. FIG. 3 is a detailed functional block diagram illustrating in more detail the configuration of a determination unit illustrated in FIG. 2. It is a flowchart explaining the basic usage method and operation | movement of a cause determination system. It is a flowchart explaining the basic usage method and operation | movement of a cause determination system. This is a result of actually calculating the Mahalanobis distance by changing the number of terms and the number of cases of the waveform vector. It is a graph which shows an experimental result. It is a graph which shows the digital signal recorded on the DAT recorder. It is a graph which shows the digital signal memorize | stored in the data storage part. It is a graph which shows the digital signal cut out by the cut-out part. It is a graph which shows the FFT process result by a frequency analysis part.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Cause determination system T Copy machine 1 Macrophone 2 DAT recorder 3 Computer main body 4a Keyboard 4b Mouse 5 Display part 3G Vector generation part 3J Judgment part 3T Time waveform vector generation part 3F Frequency waveform vector generation part 30 Data storage part 31 Audition correction part 32 Cutout unit 33 Standardization unit 34 Window function processing unit 35 Frequency analysis unit 36 Total waveform vector generation unit 37 Mahalanobis distance calculation unit 38 Final determination unit 39 Vector division unit 40 Divided vector distance calculation unit 41 Total distance calculation unit

Claims (9)

A vector generation unit that generates a vector including a plurality of data generated by generating a plurality of data based on the collected plurality of data; and
Based on a plurality of data collected a plurality of times at a stage where the cause is known, a reference space for each cause is obtained based on a plurality of known vectors generated by the vector generation unit, and a plurality of data collected in a state where the cause is unknown A determination unit that obtains a statistical distance in the reference space of the determination target vector generated by the vector generation unit based on the data of the reference and determines a cause corresponding to the determination target vector based on the statistical distance In the cause determination device comprising
The determination unit
A vector dividing unit that divides the vector generated by the vector generating unit into a plurality of small vectors each of which is an element when the elements of the vector are divided into a plurality of groups;
Based on a plurality of known small vectors obtained by dividing a plurality of known vectors by the vector dividing unit, for each of a plurality of known small vectors whose elements are mutually corresponding groups constituting a plurality of known vectors of the same cause One first reference space is obtained, and for each of a plurality of determination target small vectors obtained by dividing the determination target vector by the vector dividing unit, on each first reference space corresponding to each determination target small vector. A divided vector distance calculation unit for determining a determination target distance vector having a plurality of statistical distances corresponding to the plurality of determination target small vectors as elements, by calculating each statistical distance;
For each of the plurality of known small vectors, a second reference space is obtained based on a plurality of known distance vectors whose elements are statistical distances on each first reference space corresponding to each known small vector, An overall distance calculation unit for obtaining a statistical distance on the second reference space of the determination target distance vector obtained by the divided vector distance calculation unit;
A cause determination apparatus, comprising: a final determination unit that determines a cause corresponding to the determination target vector based on a statistical distance in the second reference space obtained by the total distance calculation unit.   The cause determination apparatus according to claim 1, wherein the vector generation unit generates a vector based on a plurality of data arranged in a time series obtained by collecting operating sounds of the target apparatus.   The cause determination device according to claim 2, wherein the determination unit is configured to determine a cause of an abnormal sound of the target device.   4. The cause determination apparatus according to claim 2, wherein the target apparatus is an image forming apparatus that conveys a recording medium and forms an image on a recording medium that is being conveyed.   The vector generation unit includes a signal cutout unit that detects predetermined reference data from a plurality of data arranged in time series and cuts out a plurality of data including data temporally preceding the reference data. The cause determination device according to claim 2, wherein the cause determination device is provided.   3. The cause determination apparatus according to claim 2, wherein the vector generation unit generates a time waveform vector having a plurality of data representing a time waveform as elements based on a plurality of data arranged in time series.   The vector generation unit generates a frequency waveform vector having a plurality of data representing a frequency distribution as elements by converting a plurality of data arranged in time series into a plurality of data in a frequency domain. The cause determination apparatus according to claim 2.   The vector generation unit is obtained by converting a plurality of data representing a time waveform obtained based on a plurality of data arranged in a time series and a plurality of data arranged in the time series into a plurality of data in a frequency domain. The cause determination apparatus according to claim 2, wherein a comprehensive waveform vector including both of the plurality of data in the frequency domain to be generated is generated.   The cause determination apparatus according to claim 1, wherein the division distance calculation unit and the total distance calculation unit calculate a Mahalanobis distance as the statistical distance.

Images