JP6236282B2 - Abnormality detection apparatus, abnormality detection method, and computer-readable storage medium - Google Patents

Description

  The present invention relates to an anomaly detection device, an anomaly detection method, and a computer-readable storage medium. For example, the present invention relates to an anomaly detection technique for detecting an anomaly in this signal from a signal after digital conversion by an A / D converter.

  From the viewpoint of maintenance of the apparatus, there is a strong demand for an abnormality detection technique for detecting wear of gears in the operating section of the apparatus and malfunction of the driving section. In particular, the operating unit in the apparatus is likely to abnormally change due to physical quantities such as sound, vibration, and distortion, and thus an abnormality is often detected using a sensor such as a microphone, a vibration sensor, or a strain sensor. However, even in these sensors, even if the sensor is provided in the vicinity of the target operating unit that performs abnormality detection, signals other than the target operating unit are mixed, and the target signal has a sufficient signal-noise ratio (S / N ratio) is often difficult to detect.

  Therefore, a signal separation technique for extracting only the signal of the target operating part from the obtained sensor signals is required. When there are a plurality of target operating parts, it is necessary to separately extract the signals of each operating part from the obtained sensor signals.

  Until now, as a signal separation technique, a separation technique using a plurality of sensor signals has been mainly studied. These techniques are techniques for separating signals using the fact that the time difference and amplitude difference between sensors differ for each signal source. As a typical signal source separation technique, there is an adaptive filter method such as a minimum dispersion beam former method. In this adaptive filter method, it is possible to separate sounds using a multi-input type filter that passes only the signal of the target operating part and does not pass other signals. In principle, the adaptive filter method can suppress a signal from a signal source having the number of sensor elements minus one.

  Therefore, in the case where there is only one signal source of the number of sensor elements in addition to the operation part to be monitored for abnormality detection as a signal source, in principle, a signal from the operation part to be monitored can be extracted with high accuracy. It becomes possible.

  On the other hand, when there are more signal sources than the number of sensor elements minus one, it is known that the accuracy of signal extraction from the monitoring target operating unit deteriorates. However, when viewed over a long period of time, even if there are more signal sources than the number of sensor elements minus one, signals are not always output from all signal sources. For example, when the signal source is an operating unit, the signal source moves or stops depending on the operation mode of the apparatus, so that a signal is output from the signal source only when it is moving.

  Therefore, even when there are more signal sources than the number of sensor elements minus one when viewed in a long time, there are many cases where there are only signal sources equal to or less than the number of sensor elements minus one when viewed in a short time. it is conceivable that. By using this, the conventional adaptive filter method aims to efficiently suppress the signal source at each time, and changes the shape of the multi-input type filter according to the sensor signal obtained from time to time. (See Non-Patent Document 1). As described above, as a technique for extracting a signal to be monitored, noise suppression using an adaptive filter that changes the value of the filter so as to effectively suppress noise at each time according to changes in the sound field environment. The law is common.

L.J.Griffith and C.W.Jim, `` An alternative approach to linearly constrained adaptive beamforming, '' IEEE Trans.Anntenas Propagation, vol.30, i.1, pp.27-34, Jan.-1982.

  However, in the configuration disclosed in Non-Patent Document 1, the follow-up performance deteriorates as the shape changes more rapidly. If the signal source type does not change for at least several seconds at a sampling rate of about 10 kHz due to sound or vibration, it can be followed, but if the signal source type changes at an extremely short time interval of less than 1 second. It becomes difficult to follow. Since the adaptive filter has a time of several seconds for learning the information of noise to be eliminated, the tracking of the operating sound of the measuring device whose sound field environment changes within a short time of less than 1 second is not in time, and works well. Noise cannot be suppressed. The same applies to a measurement apparatus that is controlled to repeatedly perform the same operation at a constant period.

  The present invention has been made in view of such a situation, and in a measurement device controlled to repeatedly perform the same operation at a constant cycle, the sound field environment changes in a short time (for example, within one second). The present invention provides a technique for effectively suppressing noise with respect to operating sound of a measurement device and extracting a signal to be monitored.

  In order to solve the above-mentioned problem, the inventors of the measurement device controlled to repeatedly perform the same operation at a constant cycle, the volume for each frequency of the operating sound is the same pattern with a constant multiple of the control cycle. We paid attention to the fact that it is a periodic sound that repeats and that the surrounding noise other than the operating sound is an aperiodic sound that exists regardless of the control period. In the present invention, it is assumed that a signal in which a periodic sound and a non-periodic sound that are repeated at a constant multiple of the control period are mixed in the sensor signal, and each periodic sound and aperiodic sound are separated. And the component which reaches | attains a sensor through the same transmission process as the operation sound of monitoring object is extracted among the separated periodic sounds, and abnormality detection is performed with respect to the extracted component.

  That is, the abnormality detection device according to the present invention executes a learning process and an abnormality detection process. The learning process uses information on the control cycle for sensor signals in which a periodic component whose volume changes at a constant multiple of the control cycle of the measuring device and a non-periodic component that varies independently of the control cycle are mixed. In this process, the maximum likelihood estimation is performed on the sensor signal to separate the periodic component and the non-periodic component, and to calculate a probability distribution that stochastically indicates the fluctuation of the feature amount of the separated periodic component. . In addition, the abnormality detection process is a process for detecting a signal abnormality using a result of the learning process with a sensor signal other than the sensor signal used in the learning process as an inspection target.

  Further features related to the present invention will become apparent from the description of the present specification and the accompanying drawings. The embodiments of the present invention can be achieved and realized by elements and combinations of various elements and the following detailed description and appended claims.

  It should be understood that the description herein is merely exemplary and is not intended to limit the scope of the claims or the application of the invention in any way.

  ADVANTAGE OF THE INVENTION According to this invention, it becomes possible to extract the signal of the operation part of monitoring object with high precision from the sensor signal which the signal component from various operation parts and the noise component which arrives from the apparatus exterior mixed.

It is a figure which shows the hardware structural example of the measuring device with which the abnormality detection apparatus of this invention is installed. It is a block diagram which shows schematic structure of the abnormality detection apparatus by embodiment of this invention. It is the figure which illustrated typically the component for every time frequency of the signal recorded with the measurement sensor 104 of the measuring device. It is a figure which shows the structural example of the display screen 1400 by embodiment of this invention. It is a figure which shows the structural example of an initial setting screen. It is a flowchart for demonstrating the outline | summary of the whole process by the abnormality detection apparatus by embodiment of this invention. It is a figure which shows the processing structure of the learning program 400 which the abnormality detection apparatus by embodiment of this invention performs. It is a figure which shows the example of estimation of the volume of each periodic component and an aperiodic component. It is a figure which shows the data structure of periodic component statistics DB404. It is a figure which shows the data structure of the transfer function of each monitoring object operation part hold | maintained by registration component DB407. It is a figure which shows the data table structure of the signal from the monitoring object operation part written in normal component DB405. It is a figure which shows the detail of the periodic component and aperiodic component learning part 403 by embodiment of this invention. 10 is a flowchart for explaining details of processing of a frequency-based periodic component / non-periodic component learning unit 802. It is a figure which shows the processing structure of the abnormality detection program 600 which the abnormality detection apparatus by embodiment of this invention performs. It is a figure which shows the process structure of the abnormality detection process 501 which the abnormality detection apparatus by embodiment of this invention performs. It is a figure which shows the detail of the periodic component / non-periodic component separation part 502 by embodiment of this invention. It is a figure which shows the modification of the periodic component / non-periodic component separation part 502 by embodiment of this invention. 10 is a flowchart for explaining details of processing of a periodic component / non-periodic component separation unit 902 for each frequency. 5 is a flowchart for explaining details of processing of an abnormality determination unit 503.

  Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, functionally identical elements may be denoted by the same numbers. The attached drawings show specific embodiments and implementation examples based on the principle of the present invention, but these are for understanding the present invention and are not intended to limit the present invention. Not used.

  This embodiment has been described in sufficient detail for those skilled in the art to practice the present invention, but other implementations and configurations are possible without departing from the scope and spirit of the technical idea of the present invention. It is necessary to understand that the configuration and structure can be changed and various elements can be replaced. Therefore, the following description should not be interpreted as being limited to this.

  Furthermore, as will be described later, the embodiment of the present invention may be implemented by software running on a general-purpose computer, or may be implemented by dedicated hardware or a combination of software and hardware.

  In the following description, the information of the present invention will be described in a “table” format. However, the information does not necessarily have to be represented by a data structure of a table, and may be a data structure such as a list, a DB, a queue, or the like. It may be expressed. Therefore, “table”, “list”, “DB”, “queue”, etc. may be simply referred to as “information” to indicate that they do not depend on the data structure.

  In the following, each process in the embodiment of the present invention will be described with “central processing unit” (also referred to as a computer or a processor) as the subject (operation subject), but various “programs” may be described as the operation subject. A part or all of the program may be realized by dedicated hardware or may be modularized. Various programs may be installed in each computer by a program distribution server or a storage medium.

<Configuration of measuring device>
FIG. 1 is a diagram illustrating a hardware configuration example of a measurement apparatus in which an abnormality detection apparatus according to an embodiment of the present invention is installed.

  The measuring apparatus 100 includes a rotary table 101, an operating unit 102 including a plurality of operating units, a monitoring target operating unit 103, and a measurement sensor 104.

  The rotary table 101 operates in accordance with the control cycle of the measuring device, and repeatedly executes the same operation at a constant cycle that is a constant multiple of the control cycle of the device. Similarly, the operating unit 102 operates at a constant cycle. For example, the piston motion is performed at a constant cycle that is a constant multiple of the control cycle. Similarly, the monitoring target operating unit 103 operates at a constant cycle that is a constant multiple of the control cycle of the apparatus.

  In the present invention, a state in which the monitoring target operation unit 103 is operating differently for some reason is detected from the sensor signal measured by the measurement sensor 104.

<Configuration of abnormality detection device>
FIG. 2 is a hardware block diagram illustrating a schematic configuration of the measurement device and the abnormality detection system (abnormality detection device + monitoring server).

  The measurement sensor 104 is provided inside the measurement apparatus 100 and is configured by a microphone, a vibration sensor, and / or a strain sensor. The analog signal captured by the measurement sensor 104 is converted from an analog signal to a digital signal by the A / D converter 201.

  The converted digital signal is sent to the abnormality detection device 200. The abnormality detection device 200 includes a central processing unit 201, a nonvolatile memory 202, and a volatile memory 203. The converted digital signal is sent to the central processing unit 201, and a learning program and an abnormality detection program described later are executed. This program is stored in the non-volatile memory 202 and is read when the program is executed. The work memory required when executing the program is secured on the volatile memory 203.

  An output value output after being processed by the abnormality detection program is sent to the HUB 206 and converted into a packet for network transmission. The converted packet is sent to the HUB 212 via the network 207. The HUB 212 transmits the received data to the monitoring server 211.

  The monitoring server 211 includes a central processing unit 208, a volatile memory 209, a volatile memory 210, a display 213, and an input device 214 such as a mouse or a keyboard. The central processing unit 208 operates the monitoring result viewer program to generate a screen to be displayed on the display 213 using the sent data. In addition, the user can control (instruct) the start of the initial setting and the start of the abnormality detection operation using the input device 214 including a mouse and a keyboard. The monitoring result viewer program is stored on the non-volatile memory 209 and is read when the program is executed. A work memory required for executing the program is secured on the volatile memory 210.

<Signals recorded by measurement sensor (example)>
FIG. 3 is a diagram schematically showing components (examples) for each time frequency of the signal recorded by the measurement sensor 104 in the measurement apparatus shown in FIG.

  The sensor signal is subjected to short-time Fourier transform and wavelet transform, which are well known to those skilled in the art, and are converted into signals for each time (τ) and frequency. In FIG. 3, the time frequency component in which the volume for each time frequency is equal to or greater than a certain threshold value is represented by a black cell. On the other hand, the time frequency component that is below a certain threshold is represented by a white cell. In FIG. 3, it can be seen that black and white cells have the same pattern in the same cycle as the control cycle.

  As shown in FIG. 3, the sensor signal measured in the measuring device having a plurality of operating parts that operate at a constant control cycle is considered to show a similar time-frequency component pattern at a cycle that is a constant multiple of the control cycle. It is done.

<Display screen example on the monitoring server>
FIG. 4 is a diagram illustrating a display example of the display screen 1400 on the display 213 on the monitoring server 211 operated by the input device 214.

  The display screen 1400 includes a plurality of buttons (initial setting button 1401, operation start button 1403, learning start button 1404, end button 1405) operated by the input device 214, and a monitoring target operating unit that displays the status of each monitoring target operating unit. A state monitoring screen 1402 as a screen configuration.

  When an initial setting button 1401 is pressed with the input device 214, an initial setting screen is displayed.

  FIG. 5 is a diagram showing a display example of the initial setting screen 1900. The initial setting screen 1900 displays a screen on which the user can set a plurality of buttons and sensor strength (bj (f) described later) of the monitoring target operating unit. Specifically, the user sets an index of the monitoring target operating unit for setting the sensor strength on the initial setting screen 1900, sets the strength of each sensor, and then presses a registration button 1901. Accordingly, bj (f) of the corresponding monitoring target operating unit can be registered in the registered component DB 407 described later. Normally, one measurement sensor is installed in the vicinity of each monitoring target operating unit. For this reason, it is assumed that the intensity of the measurement sensor closest to the monitoring target working unit is set to 1 and the intensity of the other measurement sensors is set to 0. The user presses an initial setting end button 1902 after setting the information of the monitoring target operating unit. As a result, the initial setting screen 1900 is terminated and the display screen 1400 is restored.

  Referring to FIG. 4 again, on the display screen 1400, when the user presses the learning start button 1404, a learning program 400 described later is executed on the central processing unit 201. Further, when the user presses the operation start button 1403, the abnormality detection program 501 is executed. Further, when the user presses an end button 1405, the viewer program ends.

  The monitoring target operating unit status monitoring screen 1402 displays whether each monitoring target operating unit is normal or abnormal.

<Overall processing in the abnormality detection device>
FIG. 6 is a diagram showing a typical operation flow of each program in the abnormality detection system. The operation flow starts when the user gives an instruction using the input device 214 on the display screen 1400 shown in FIG.

  First, the central processing unit 208 starts a viewer program for displaying the display screen 1400 on the monitoring server 211 (S701).

  Next, the central processing unit 201 executes the learning program 400 in response to the learning start button 1404 being pressed by the user (S702). By executing this learning program, the periodic component and the non-periodic line segment of the input signal are separated, and only the signal (normal component) of the monitoring target movable part is registered in the database (normal component DB 405).

  Thereafter, the central processing unit 201 operates the abnormality detection program 600 in response to the operation start button 1403 being pressed by the user (S703). In the abnormality detection program 600, an abnormality detection process 501 (see FIG. 15) is executed every time a sensor signal is obtained.

  The central processing unit 208 of the monitoring server 211 confirms whether the viewer program end button has been pressed each time the abnormality determination result of the abnormality detection processing 501 is received by the monitoring server 211 (S704).

  If the button is pressed (Yes in S704), the central processing unit 208 ends the viewer program and the central processing unit 201 ends the abnormality detection program 600 (S705).

<Learning program>
FIG. 7 is a diagram showing a processing block configuration of the learning program 400 executed by the central processing unit 201.
The recording unit 401 buffers the sensor signal recorded by the measurement sensor 104. Here, the measured sensor signal is expressed as shown in Equation (1).

ここで、Mは計測センサの数とし、tはサンプリングの度にインクリメントされるＡ／Ｄ変換開始後からのサンプリング数とする。Tは行列またはベクトルの転置演算子とする。 [Formula 1]

Here, M is the number of measurement sensors, and t is the number of samplings after the start of A / D conversion that is incremented every sampling. T is a matrix or vector transpose operator.

  The frequency resolving unit 402 converts the time-domain signal of each measurement sensor into a time-frequency domain by a known frequency resolving process such as short-time Fourier transform or wavelet transform. Here, if m is an index indicating the number of the measurement sensor, the time domain signal of the m-th measurement sensor included in x (t) is xm (t). A signal obtained by converting the time domain signal into the time frequency domain is expressed as xm (f, τ). Here, f is an index representing a frequency, and τ is a variable representing a frame number incremented every time a short-time Fourier transform is performed. f is an integer value from 1 to K.

Here, the sensor signal measured in the time frequency domain is expressed as shown in Equation (2).
[Formula 2]

  The periodic component / non-periodic component learning unit 403 first stores x (f, τ) of a plurality of frames in a buffer. Here, it is assumed that signals up to τ = 1, 2,..., LLt are stored in the buffer. LLt is a natural number times the control cycle of the measuring device. Then, the periodic component / non-periodic component learning unit 403 decomposes x (f, τ) into periodic components / non-periodic components according to a processing flow described later. In the decomposition, the number of periodic components and the number of non-periodic components are assumed to be given in advance. Let the periodic component after decomposition be ci (f, τ) (i is the index of the periodic component), and the non-periodic component be rj (f, τ) (j is the index of the aperiodic component). Regarding the volume of the periodic component ci (f, τ) after decomposition, when the average volume averaged in the time frequency direction is below a certain ratio compared to the average volume averaged in the time frequency direction of x (f, τ) It is evaluated that the periodic component does not exist. By performing this process, even if the number of periodic components actually present is smaller than the number of periodic components set in advance, it is possible to extract periodic components that match the actual number of periodic components.

  The periodic component / non-periodic component learning unit 403 separates the extracted periodic component ci (f, τ) and associates it (corresponds), and the volume vi (f, τ) of the corresponding periodic component for each time frequency, and A time-invariant covariance matrix Ri (f) and a value (Pi) obtained by dividing the period of the extracted periodic component by the control period are output. Also, the volume for each time frequency is output for the non-periodic component.

  Here, estimation of the volume of the periodic component and the non-periodic component will be described. FIG. 8 shows an example of estimating the volume of each periodic component and aperiodic component. The periodic component is estimated as a component whose volume changes in a natural number times the control period. On the other hand, the volume of the non-periodic component is estimated as a component having no period depending on the control period.

  Returning to FIG. 7, assuming that the number of extracted periodic components is N, the periodic component statistics DB 404 stores N periodic components Pi, volume vi (f, τ) for each time-frequency, and common for each frequency. The variance matrix Ri (f) is stored. FIG. 9 shows the data structure of the periodic component statistics DB 404. Assuming that the control period of the device is Lt, for each periodic component, the signal period / control period, the total frequency (1 to K), and the time-frequency for the number of frames corresponding to the signal period (Pi × Lt) of each periodic component The volume and the covariance matrix of all frequencies are stored in the periodic component statistics DB 404. Note that the volume vi (f, τ) for each time-frequency is information on the intensity that changes with time, and the covariance matrix Ri (f) for each frequency is converted into an intensity ratio that takes a constant value regardless of time. Information.

  Returning to FIG. 7, the registration determination processing 406 determines whether or not the extracted periodic component is the operation sound to be monitored. If the extracted periodic component is a component of the monitoring target operation unit, the transfer function and probability distribution of the periodic component are obtained. Obtained from the registered component DB 407 and written to the normal component DB 405. Here, whether or not the component is a component of the monitoring target operating unit is determined in advance (at the time of initial setting) by the transfer function of the component of the monitoring target operating unit held in the registered component DB 407 and the time-invariant covariance matrix Ri of the periodic component. Estimated from the degree of agreement with (f).

<Transfer function of monitoring target operating part>
FIG. 10 shows the data structure of the transfer function of each monitoring target operating unit held in the registered component DB 407. As illustrated in FIG. 10, the transfer function bj (f) is registered in the registered component DB 407 for each monitoring target operating unit. If Lm is the number of monitoring target operating parts, bj (f) is a transfer function of the frequency f of the jth monitoring target operating part, and includes elements corresponding to the number of sensor elements. The degree of coincidence with Ri (f) is determined using the following equation (3).

ここで、traceは行列のトレースを算出する演算子である。Hは行列またはベクトルのエルミート転置を取る演算子とする。モニタリング対象稼動部ごとに（jごとに）式（３）が最大となるような周期成分を選択し、その周期成分がj番目のモニタリング対象稼動部の周期成分として正常成分ＤＢ４０５に書き込まれる。 [Formula 3]

Here, trace is an operator for calculating a matrix trace. Let H be an operator that takes a Hermitian transpose of a matrix or vector. A periodic component that maximizes Equation (3) is selected for each monitoring target operating unit (for each j), and the periodic component is written in the normal component DB 405 as a periodic component of the jth monitoring target operating unit.

<Normal component DB>
FIG. 11 is a diagram illustrating a configuration of a data table of signals from the monitoring target operating unit to be written in the normal component DB 405. In the table, a transfer function aj (f) is registered for each monitoring target operating unit.

  The transfer function aj (f) may be registered as bj (f) or may be registered as the first eigenvector of the covariance matrix Ri (f) selected as the periodic component of the j-th monitoring target operating unit. Good. By registering it as the first eigenvector of Ri (f), it is possible to register transfer functions that include effects that cannot be known in advance by physical simulation at the time of device design, such as the effects of reverberation and reverberation in the measurement device. It can be expected that the determination accuracy of the abnormality determination process is improved.

  In the normal component DB 405, the periodic component probability distribution pj (x) is registered. This probability distribution pj (x) is information indicating the normal range of the normal component signal from the modeling target working unit, and abnormality determination is performed using this information. Here, the probability distribution of the periodic component is a model that stochastically represents the fluctuation of the feature amount of the periodic component, and x is an acoustic feature amount. As the probability distribution, various distributions conventionally used as an abnormality detection method such as a mixed normal distribution well known to those skilled in the art can be used. Also, the MFCC (Mel Frequency Cepstrum Coefficients) obtained from the separation signal ci (f, τ) of each periodic component, ci (f, τ) of all frequencies and all frames, and the volume thereof are vectorized to obtain the acoustic feature amount x. Used as In other words, by extracting the component that is a feature of the sound from the periodic component ci (f, τ) of the input signal and checking the distribution of the component that is the feature of the sound, the probability distribution pj ( x) can be obtained.

  Further, in order to obtain a probability distribution such as GMM (Gaussian Mixture Model) from data, a plurality of measurements are required for calculating a statistic. In this case, the abnormality detection program 501 is executed a plurality of times, and a periodic component detected as a signal of the j-th monitoring target operating unit is regarded as a plurality of times of measurement, and a probability distribution of GMM is obtained from the data.

  The transfer function bj (f) may be configured to be set by a user interface for initial setting (see FIG. 5). Further, a transfer function when a signal is transmitted from the monitoring target operating unit to the sensor may be generated from the relative positional relationship between the monitoring target operating unit and the sensor, and the transfer function may be given as bj (f). For example, when the measurement sensor is a microphone, the transfer function when the signal is transmitted from the monitoring target operating unit to the sensor is attenuated in proportion to the square of the distance between the monitoring target operating unit and the sensor. By dividing by the speed of sound, it is possible to adopt a configuration in which bj (f) is obtained from the time delay that can be obtained. There are many methods known to those skilled in the art for obtaining a transfer function from vibrations and strains by physical simulation from design information, but bj (f) is widely applied to those methods. Can be obtained. It is assumed that the magnitude of bj (f) is adjusted so that the magnitude (square norm) of vector bj (f) is 1.

<Internal configuration of periodic component / non-periodic component learning unit>
FIG. 12 is a diagram showing an internal block configuration of the periodic component / non-periodic component learning unit 403.
The periodic component / non-periodic component learning unit 403 includes a plurality of frequency-specific periodic component / non-periodic component learning units 802-1 to 802-1 to K and a permutation resolution unit 803.

  The frequency periodic component / non-periodic component learning unit 802-1... K separates the sensor signal x (f, τ) measured in the time-frequency domain into a periodic component and an aperiodic component for each frequency f. The sensor signal varies with time for each frequency and has a different period for each frequency (see FIG. 3). Therefore, the sensor signal is divided into a plurality of bands for learning. Learning the periodic component and the non-periodic component for each frequency is one of the features of the present invention.

  In the separation process, it is impossible to design in advance which index the periodic component or the non-periodic component will be, and take a random value for each frequency. For example, there is a possibility that the first periodic component of the index at frequency 1 and the first periodic component of the index at frequency 2 are signals from completely different signal sources.

  Therefore, the permutation resolution unit 803 performs a process of changing the index so that the signals from the same signal source are generated if the indexes of the separated frequency components are the same. Since a plurality of frequency components are output from the same operating part, it is necessary to associate each component with the operating part. Permutation resolution processing is known and can be performed by those skilled in the art, but for example, processing for indexing is performed so that the correlation in the frame direction of volume vi (f, τ) for each frame is high. (Each frequency component output from the same operating unit is correlated because it has a high correlation), so that appropriate index reassignment can be performed. The output signal of the permutation resolution unit 803 can be expected to be a signal from the same signal source as long as the index is the same even at different frequencies.

<Details of frequency component / aperiodic component learning processing for each frequency>
FIG. 13 is a flowchart for explaining details of processing executed by the frequency-periodic component / non-periodic component learning unit 802. In order to separate a periodic component and a non-periodic component from an input signal, parameters for separation (for example, information on intensity and information on which frequency component becomes larger in which time zone) are required. However, to obtain the parameters, the input signal must first be separated. From the separated signal, it is learned that a certain frequency component increases / decreases in a certain time zone. Since such learning cannot be completed by a single process, the process shown in FIG. 13 is executed.

  The central processing unit 201 starts a process of learning each periodic component and non-periodic component from the sensor signal x (f, τ) measured in the time-frequency domain (S1101).

  Further, the central processing unit 201 sets provisional parameters (S1102). In the temporary parameter setting process, an appropriate parameter (random) is first set. Since it is a random parameter, the accuracy is not good, but it can be said that it is improved compared to the input signal. More specifically, first, the numbers of preset periodic components and aperiodic components are read as set values. The period of each periodic component is an integral multiple of the control period, and the period is read from the set value even in the period of each periodic component. Each periodic component and non-periodic component are represented by the same index i. Further, the volume vi (f, τ) of the time frequency of each periodic component and aperiodic component is set to 1, and the covariance matrix Ri (f, τ) is given as a random Hermitian matrix. Here, Ri (f, τ) is a matrix of M rows and M columns. Also, vi (f, τ) is a scalar value.

  Then, the central processing unit 201 separates and extracts each periodic component and non-periodic component from vi (f, τ) and Ri (f, τ) (S1103). Here, τ is from 1 to LLt. Separation extraction is performed using a multi-channel Wiener filter. Specifically, the separated signal ci (f, τ) is obtained by ci (f, τ) = wi (f, τ) x (f, τ). Ri (f, τ) is expressed by equation (4), and Wiener filter wi (f, τ) is expressed by equation (5). Furthermore, in S1103, the central processing unit 201 calculates a square error by the equation (6). In Equation (6), I is an M × M unit matrix.

［式５］

［式６］

[Formula 4]

[Formula 5]

[Formula 6]

  Next, the central processing unit 201 updates the separation parameter update using Expression (7) (S1104). Expression (7) is an expression for maximum likelihood estimation. That is, in the present invention, one of the features is that maximum likelihood estimation is performed using periodicity information.

[Formula 7]

  Furthermore, in S1104, when the i-th signal is a periodic component and the period is Lt, the central processing unit 201 transforms vi (f, τ). Specifically, first, a set T (r) of τ that satisfies τ = Lt * n + r (n is an arbitrary natural number) is obtained for each divided time r in one cycle. Then, for τ included in T (r), an average value of vi (f, τ) is taken and is set as vi (f, r). Next, vi (f, τ) = vi (f, r) is set for τ included in T (r). Further, the covariance matrix Ri (f, τ) is obtained (updated) using Equation (8).

[Formula 8]

  Then, the central processing unit 201 checks whether the number of executions of the separation parameter update has reached a predetermined number of times, or whether the square error calculated by the equation (6) is smaller than a predetermined value. When the number of times has been reached, or when the square error is small, it is determined that the learning has ended (S1105). If it is determined to end, the process proceeds to learning end setting (S1106), and the learning ends. If it is not determined to end, the process proceeds to separation process execution (S1103).

  As described above, in the learning process, separation accuracy is improved by alternately performing separation of periodic and aperiodic components and parameter update.

<Abnormality detection program configuration>
FIG. 14 is a diagram showing a block configuration of the abnormality detection program 600. The abnormality detection program includes an abnormality detection process 501 for detecting an abnormality and a detection result transmission process 601 for sending the detected abnormality to the monitoring server 211.

  FIG. 15 is a diagram illustrating a detailed block configuration of the abnormality detection processing 501. The abnormality detection process 501 has the same configuration as the learning program 400 shown in FIG. In the abnormality detection process 501, LLt is not necessarily set to the same value as that of the learning program 400, but is set to a natural number times the control cycle of the measuring apparatus.

  First, the periodic component / non-periodic component separation unit 502 obtains the periodic component statistics registered in the periodic component statistics DB 404 (periodic components in all frequency bands are registered in association with the monitoring target operating unit). To separate the periodic and aperiodic components. If the average volume averaged in the time frequency direction of the decomposed periodic component ci (t) is less than a certain ratio compared to the average volume averaged in the time frequency direction x (t), the periodic component Is considered nonexistent. By performing this process, even if the number of periodic components actually present is smaller than the number of periodic components set in advance, it is possible to extract periodic components that match the actual number of periodic components.

  The abnormality determination unit 503 determines whether or not the extracted periodic component includes an abnormality, using information on the monitoring target operation unit registered in the normal component DB 405. That is, the deviation degree of the extracted periodic component is checked based on the probability distribution pj (x) of the monitoring target operating unit, and abnormality determination is executed depending on whether or not the deviation degree is larger than a predetermined value.

<Detailed configuration of periodic component / non-periodic component separation unit>
FIG. 16 is a diagram illustrating a detailed block configuration of the periodic component / non-periodic component separation unit 502 according to the embodiment of the present invention.

  Periodic component / non-periodic component separating units 902-1 to 902-1 for each frequency separate the input signal x (f, τ) into periodic components / non-periodic components for each frequency. The per-frequency periodic component / non-periodic component separators 902-1 to 902-1 are configured by filters. Since the information of unnecessary signal intensity (periodic component) in each time zone is acquired from the periodic component statistics DB 404, a filter capable of removing the unnecessary signal is provided, and the filter is used to monitor from the monitoring target operating unit. Get only the signal. Since non-periodic components vary depending on time and place, they are separated while updating parameters in real time (see FIG. 18).

  The inverse frequency transform unit 903 transforms each periodic component ci (f, τ) into a time domain signal ci (t) by inverse Fourier transform or inverse wavelet transform. The converted time domain signal ci (t) is output. However, the inverse frequency conversion unit 903 is not an essential configuration.

  Note that the periodic component / non-periodic component separation unit 502 can also have the same configuration as the periodic component / non-periodic component learning unit 403. FIG. 17 is a diagram showing such a configuration, which makes it possible to separate signals in real time while learning. Since the processing flow has already been described, the description thereof is omitted.

<Details of frequency / periodic component separation for each frequency>
FIG. 18 is a flowchart for explaining details of processing by the frequency-periodic component / non-periodic component separation unit 902.

  The central processing unit 201 executes separation processing every time an input signal having a predetermined frame length is accumulated (S1201).

  When the separation process is started, the central processing unit 201 first reads the parameters of the periodic component from the periodic component statistics DB 404 (S1202). The periodic component parameters are vi (f, τ) and Ri (f), which are registered in the periodic component statistics DB 404.

  On the other hand, since the signal source changes with time for the non-periodic component, the central processing unit 201 sets a real-time update parameter (S1203). Here, for the non-periodic component, for example, vi (f, τ) = 1 and Ri (f) is a random Hermitian matrix.

  Then, the central processing unit 201 executes separation processing (S1204). Specifically, each periodic component and non-periodic component are separated and extracted from vi (f, τ) and Ri (f, τ). Here, τ is from 1 to LLt, and the separated signal ci (f, τ) is obtained by ci (f, τ) = wi (f, τ) x (f, τ) using a multi-channel Wiener filter. Ri (f, τ) is expressed by the above equation (4), and the Wiener filter wi (f, τ) is expressed by the above equation (5). Furthermore, in S1204, the central processing unit 201 calculates a square error by the above-described equation (6). In Equation (6), I is an M × M unit matrix.

  Next, the central processing unit 201 updates the real-time update parameter using the above equation (7) (S1205). When the i-th signal is a periodic component and the period is Lt, the central processing unit 201 transforms vi (f, τ). Specifically, first, a set T (r) of τ that satisfies τ = Lt * n + r (n is an arbitrary natural number) is obtained for each divided time r in one cycle. Next, with respect to τ included in T (r), an average value of vi (f, τ) is taken and is defined as vi (f, r). Then, τ included in T (r) is set to vi (f, τ) = vi (f, r). On the other hand, for only the aperiodic component, the covariance matrix Ri (f, τ) is obtained (updated) using the above-described equation (8).

  Subsequently, the central processing unit 201 performs an update end determination (S1206). For example, if the number of times the real-time update parameter has been updated has reached a predetermined number of times, it is determined that the process has ended. If it is determined that the process is to be terminated (Yes in S1206), the process proceeds to S1207 (setting for separation termination), and the separation process is terminated. If it is not determined that the process has been completed (No in S1205), the process returns to S1204 (separate process execution). The end may be determined depending on whether the calculated square error is smaller than a predetermined value.

<Abnormality judgment processing>
FIG. 19 is a flowchart for explaining details of processing by the abnormality determination unit 503.
The central processing unit 201 starts abnormality determination each time the periodic component and non-periodic component separation processing 502 ends (S1301).

  Next, the central processing unit 201 sets the periodic signal source index for performing the determination process to 1 (S1302). Therefore, the abnormality determination process is executed from the periodic component having the smallest index among the separated periodic components.

  The central processing unit 201 extracts the transfer function of the periodic component to be processed as the first eigenvalue di (f) of Ri (f) (S1303).

  The central processing unit 201 then calculates the distance between the transfer function aj (f) of the monitoring target operating unit registered in the normal component DB 405 and the transfer function (first eigenvalue di (f)) of the periodic component to be processed as Calculation is performed using (9), and j having the smallest calculated distance is selected (S1304).

[Formula 9]

  The central processing unit 201 calculates the proximity of the signal source according to the equation (10), and determines whether or not it is equal to or less than a predetermined threshold TH1 (S1305). If the proximity of the signal source is equal to or smaller than the threshold TH1 (Yes in S1305), the process proceeds to S1306. If greater than TH1 (No in S1306), the process proceeds to S1309.

[Formula 10]

  In S1306, the central processing unit 201 extracts the acoustic feature quantity x from the time domain signal ci (t) of the periodic component i (S1306).

  Next, the central processing unit 201 substitutes the extracted acoustic feature quantity for x of the normal component probability distribution pj (x) stored in the normal component DB 405 to calculate the likelihood pj (x) (S1307).

  Subsequently, the central processing unit 201 determines whether the calculated likelihood is equal to or greater than a predetermined threshold TH2 (S1308). If the likelihood is greater than or equal to the threshold value TH2 (Yes in S1308), the process proceeds to S1309. If the likelihood is smaller than TH2 (No in S1308), the process proceeds to S1311.

  In S1309, the central processing unit 201 determines that the extracted periodic component to be processed is normal, and adds 1 to the periodic signal source index (S1309).

  The central processing unit 201 determines whether the periodic signal source index is equal to or less than the number of extracted periodic signal sources (S1310). If the number is equal to or less than the number of frequency sources (Yes in S1310), the process proceeds to determination of the next periodic component (S1303). Otherwise, normality is notified and the process ends (S1312).

  On the other hand, when it is determined that the likelihood calculated in S1308 is smaller than the predetermined threshold TH2, the central processing unit 201 determines that the j-th monitoring target operating unit closest to the i-th is abnormal and ends the processing ( S1311).

<Summary>
(I) According to the embodiment of the present invention, sensor signals measured by a plurality of measurement sensors provided in the measurement device are input to the abnormality detection device. In this sensor signal, a periodic component whose volume changes by a constant multiple of the control period of the measuring device and an aperiodic component that varies independently of the control period are mixed. First, the abnormality detection device performs maximum likelihood estimation on the sensor signal using the information on the control cycle of the measurement device (the frequency component cycle is a constant multiple of the control cycle) (initially suitable as a parameter). The volume vi (f, τ) and the covariance matrix Ri (f, τ) are given, and the parameter update is executed until the square error of the periodic component ci (f, τ) becomes smaller than a predetermined value or a predetermined number of times) Thus, the periodic component and the non-periodic component are separated. Then, the abnormality detection device calculates a probability distribution that stochastically indicates the fluctuation of the feature amount of the separated periodic component, and registers it in the database together with the transfer function of the monitoring target operating unit (the learning process). When there are a plurality of operating units, the acoustic signal from the monitoring operating unit is based on the preset transfer function of the monitoring operating unit and the time-invariant information of the periodic component of the acoustic signal from the monitoring operating unit. Periodic components are extracted.

  Next, the anomaly detection apparatus uses a sensor signal other than the sensor signal used in the learning process as an inspection target, and uses a result of the learning process (probability distribution information) to detect a signal greatly deviating from the learned periodic component. By detecting the presence / absence, an abnormality of the sensor signal to be inspected is detected (abnormality detection processing). In this way, the periodic component and the non-periodic component are separated from the sensor signal using maximum likelihood estimation to detect anomalies in the periodic component. It is possible to follow, and to effectively suppress noise and extract a signal to be monitored.

(Ii) In the learning process, the abnormality detection device divides the received sensor signal into a plurality of frequency bands, separates the signals in each frequency band into periodic components and aperiodic components (separation processing for each frequency band), Based on the correlation of the signals in the frequency band, the periodic component of the monitoring target working unit is extracted (permutation solution processing). By processing in this way, even a signal that varies with time for each frequency can be separated into a periodic component and an aperiodic component.

(Iii) In the embodiment of the present invention, the same process as the learning process is executed when a signal abnormality is detected. In other words, the abnormality detection apparatus performs maximum likelihood estimation on the target sensor signal (input signal) using the periodic component parameter obtained by the learning process, thereby setting the target sensor signal as the periodic component. Separate into non-periodic components. Then, the abnormality detection apparatus includes the acoustic feature amount of the periodic component obtained by separation and the probability distribution of the monitoring target operation unit calculated by the learning process (the periodic component in which the separated periodic component is registered in the normal component DB). Is detected), the abnormality of the target sensor signal is detected.

  The abnormality detection device divides the target sensor signal into a plurality of frequency bands in the abnormality detection process. Then, the abnormality detection apparatus separates the signal of each frequency band into a periodic component and an aperiodic component using a filter that removes a signal from an operating unit other than the monitoring target operating unit, and a signal of the periodic component of the monitoring target operating unit To take out. By doing in this way, it becomes possible to efficiently extract only the periodic component of the desired signal (signal from the monitoring target operating unit) as the inspection target.

(Iv) The present invention can also be realized by software program codes that implement the functions of the embodiments. In this case, a storage medium in which the program code is recorded is provided to the system or apparatus, and the computer (or CPU or MPU) of the system or apparatus reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the functions of the above-described embodiments, and the program code itself and the storage medium storing the program code constitute the present invention. As a storage medium for supplying such program code, for example, a flexible disk, CD-ROM, DVD-ROM, hard disk, optical disk, magneto-optical disk, CD-R, magnetic tape, nonvolatile memory card, ROM Etc. are used.

  Also, based on the instruction of the program code, an OS (operating system) running on the computer performs part or all of the actual processing, and the functions of the above-described embodiments are realized by the processing. May be. Further, after the program code read from the storage medium is written in the memory on the computer, the computer CPU or the like performs part or all of the actual processing based on the instruction of the program code. Thus, the functions of the above-described embodiments may be realized.

  Further, by distributing the program code of the software that realizes the functions of the embodiment via a network, it is stored in a storage means such as a hard disk or memory of a system or apparatus, or a storage medium such as a CD-RW or CD-R And the computer (or CPU or MPU) of the system or apparatus may read and execute the program code stored in the storage means or the storage medium when used.

  Finally, it should be understood that the processes and techniques described herein are not inherently related to any particular apparatus, and can be implemented by any suitable combination of components. In addition, various types of devices for general purpose can be used in accordance with the teachings described herein. It may prove useful to build a dedicated device to perform the method steps described herein. Various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiments. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined. Although the present invention has been described with reference to specific examples, these are in all respects illustrative rather than restrictive. Those skilled in the art will appreciate that there are numerous combinations of hardware, software, and firmware that are suitable for implementing the present invention. For example, the described software can be implemented in a wide range of programs or script languages such as assembler, C / C ++, perl, shell, PHP, Java (registered trademark).

  Furthermore, in the above-described embodiment, control lines and information lines are those that are considered necessary for explanation, and not all control lines and information lines on the product are necessarily shown. All the components may be connected to each other.

  In addition, other implementations of the invention will be apparent to those skilled in the art from consideration of the specification and embodiments of the invention disclosed herein. It should be noted that the specification and specific examples are merely exemplary, and the scope and spirit of the present invention are set forth in the following claims.

DESCRIPTION OF SYMBOLS 100 ... Measuring apparatus 101 ... Rotary table 102 ... Operating part 103 ... Monitoring object operating part 104 ... Measuring sensor 200 ... Abnormality detecting apparatus 201 ... Central processing unit 202 ... -Nonvolatile memory 203 ... Volatile memory 211 ... Monitoring server

Claims (15)

An abnormality detection device that detects an abnormality of a sensor signal measured by a plurality of measurement sensors included in the measurement device,
Stored is a sensor signal that is the measured sensor signal and includes a periodic component whose volume changes by a constant multiple of the control period of the measuring device and an aperiodic component that varies independently of the control period. Memory,
A processor for detecting an abnormality of the sensor signal,
The processor is
By reading the sensor signal from the memory and performing maximum likelihood estimation on the sensor signal using the control period information, the periodic component and the non-periodic component are separated, and the separated period A learning process for calculating a probability distribution that stochastically indicates fluctuations in the feature amount of the component;
A sensor signal other than the sensor signal used for the learning process is read from the memory as an inspection target, and an abnormality detection process for detecting an abnormality of the sensor signal of the inspection target using a result of the learning process;
An abnormality detection apparatus characterized by executing In claim 1,
The sensor signal includes acoustic signals from a plurality of operating units including the monitoring target operating unit that the measurement device has,
The processor, in the learning process, based on the preset transfer function of the monitoring operation unit and the time invariant information of the periodic component of the acoustic signal from the monitoring operation unit, the acoustic signal from the monitoring operation unit An abnormality detection apparatus that extracts the periodic component. In claim 2,
In the learning process, the processor divides the sensor signal into a plurality of frequency bands, separates the signal in each frequency band into the periodic component and the non-periodic component, and the correlation between the signals in each frequency band And a permutation solution process for extracting a periodic component of the monitoring target operating unit based on the abnormality detection device. In claim 2,
In the abnormality detection process, the processor performs maximum likelihood estimation on the sensor signal to be inspected using the parameter of the periodic component obtained by the learning process, so that the sensor signal to be inspected Is divided into the periodic component and the non-periodic component, and the inspection is performed based on the acoustic feature quantity of the periodic component obtained by the separation and the probability distribution of the monitoring target operating unit calculated in the learning process. An abnormality detection device for detecting an abnormality of a target sensor signal. In claim 4,
In the abnormality detection process, the processor divides the sensor signal to be inspected into a plurality of frequency bands, and uses a filter that removes a signal from an operating unit other than the monitoring target operating unit, to output a signal in each frequency band. An abnormality detection apparatus that performs processing of separating the periodic component and the non-periodic component and extracting a periodic component signal of the monitoring target operating unit. In claim 2,
The processor further performs a process of displaying a result of an abnormality detection process indicating whether the operation of the monitoring operation unit is normal or abnormal on a display screen of a display device. In claim 2,
The processor further displays an initial setting screen for setting the monitoring rig target operating unit from the plurality of operating units and setting the strengths of the plurality of measurement sensors corresponding to the monitoring target operating unit. An abnormality detecting device characterized by executing processing to be displayed on the screen. An abnormality detection method for detecting an abnormality of a sensor signal measured by a plurality of measurement sensors included in a measurement device,
A sensor signal in which the processor is a sensor signal in which a periodic component whose volume changes by a constant multiple of the control period of the measuring device and a non-periodic component which varies independently of the control period are mixed. Reading the sensor signal from a memory storing
The processor performs maximum likelihood estimation on the sensor signal using the control period information, thereby separating the periodic component and the non-periodic component, and obtaining a feature amount of the separated periodic component. A learning step for calculating a probability distribution indicating the variation stochastically;
An abnormality detection step in which the processor reads a sensor signal other than the sensor signal used in the learning step from the memory as an inspection target, and detects an abnormality of the sensor signal of the inspection target using a processing result of the learning step; ,
An abnormality detection method comprising: In claim 8,
The sensor signal includes acoustic signals from a plurality of operating units including the monitoring target operating unit that the measurement device has,
In the learning step, the processor, based on a preset transfer function of the monitoring operation unit and time-invariant information of the periodic component of the acoustic signal from the monitoring operation unit, the acoustic signal from the monitoring operation unit An abnormality detection method, wherein the periodic component is extracted. In claim 9,
In the learning step, the processor divides the sensor signal into a plurality of frequency bands, separates the signal in each frequency band into the periodic component and the non-periodic component, and the correlation between the signals in each frequency band And a permutation solving process for extracting a periodic component of the monitoring target working unit based on the abnormality detection method. In claim 9,
In the abnormality detection step, the processor performs maximum likelihood estimation on the sensor signal of the inspection target using the parameter of the periodic component obtained by the processing of the learning step, thereby Based on the acoustic feature quantity of the periodic component obtained by separating the sensor signal into the periodic component and the non-periodic component and the probability distribution of the monitoring target operating unit calculated in the learning process, An abnormality detection method comprising detecting an abnormality of a sensor signal to be inspected. In claim 11,
In the abnormality detection step, the processor divides the sensor signal to be inspected into a plurality of frequency bands, and uses a filter that removes a signal from an operating unit other than the monitoring target operating unit to output a signal in each frequency band. An abnormality detection method, wherein the periodic component and the non-periodic component are separated and a periodic component signal of the monitoring target operating unit is extracted. The claim 9, further comprising:
The abnormality detection method characterized by including the step in which the said processor displays the result of the abnormality detection process which shows whether operation | movement of the said monitoring operation | movement part is normal or abnormal on the display screen of a display apparatus. The claim 9, further comprising:
Wherein the processor, the plurality of setting the monitoring in g object operation unit from the operating unit, display an initial setting screen for setting the respective intensities of the plurality of measurement sensors in correspondence with the monitored operating unit The abnormality detection method characterized by including the step displayed on this. A computer-readable storage medium for storing a program for causing a computer to execute an abnormality detection method,
The abnormality detection method is an abnormality detection method for detecting an abnormality of a sensor signal measured by a plurality of measurement sensors included in a measurement device,
A sensor in which the computer is a sensor signal in which a periodic component whose volume changes by a constant multiple of a control period of the measuring device and a non-periodic component that varies independently of the control period are mixed Reading the sensor signal from a memory storing the signal;
The computer separates the periodic component and the non-periodic component by performing maximum likelihood estimation on the sensor signal using the information on the control cycle, and determines the feature amount of the separated periodic component. A learning step for calculating a probability distribution indicating the variation stochastically;
An abnormality detecting step in which the computer reads a sensor signal other than the sensor signal used in the learning step from the memory as an inspection target, and detects an abnormality of the sensor signal of the inspection target using a processing result of the learning step; ,
A computer-readable storage medium comprising:

Images