JP2008065393A - Group discrimination device and group discrimination method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a new group discrimination device in which an SVM(support vector machine) learns a shape of a whole spectrum as well as a certain partial feature quantity. <P>SOLUTION: A group discrimination device 1 is provided with: a spectrum estimation means 2 for estimating a discrete power spectrum from time-space series data; a normalization means 3 for forming a normalized histogram by normalizing the spectrum; a histogram storage means 4 for storing the feature vector of the normalized histogram in association with the group; a discrimination hyperplane generation means 5 for classifying time-space series data for learning into a plurality of groups by using an SVM 6 for forming a discrimination hyperplane for classifying the feature vector in a feature space by using kernel functions; and a discrimination means 7 for discriminating a group to which the target time-space series data belongs by using the discrimination hyperplane. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a group discrimination device and a group discrimination method. More particularly, the present invention relates to a group discrimination apparatus and a group discrimination method for discriminating a group to which space-time series data belongs by statistical analysis using a support vector machine.

  There are many methodologies for discriminating given data, but the development of discriminative learning machines using the kernel method represented by SVM (Support Vector Machine) supports various fields. Giving. Prior to the introduction of kernel method and machine learning, discriminating machines project given high-dimensional data into a low-dimensional space that well reflects its characteristics, and then extract a small number of parameters that are useful for discrimination. Met. For example, a robot has been proposed that acquires the name and feature data of a target object, stores the association information, learns the name of the object, and determines whether the object is a new object. This uses the acoustic features of a human voice and the morphological features of a part of the face. (See Patent Document 1)

  In addition, there is a fault diagnosis method as an application of time-series data group discrimination. This is mainly time-series analysis, frequency spectrum analysis, spatial domain analysis, or a combination of these, including frequency. Spectral analysis is often used for fault diagnosis of rotation / vibration mechanisms. In the frequency spectrum analysis method, a small number of feature amounts are extracted from the frequency peak of the mode of interest such as the half-width method, and the properties of the mode are examined from these feature amounts.

  Considering the failure diagnosis of the high-pressure gas pressure regulator, steady security of the gas supply equipment is essential due to the large number of consumers and the danger of high-pressure gas. The high-pressure gas pressure regulator is one of the gas supply devices, and the pressure in the high-pressure gas tank is reduced to a pressure that can be safely used in home gas devices by a feedback system based on the vertical movement of the rubber film installed inside it. adjust. However, it has been reported that, in rare cases, an abnormal vibration occurs inside the regulator even though it is within the expiration date. However, the rubber film inside the regulator cannot be observed directly because of its strong explosion-proof property. There is a method using a laser Doppler vibrometer to detect the failure of the regulator, but this instrument is expensive, heavy and complicated to use. Therefore, the inventors have developed a measurement system for a rubber diaphragm that is inexpensive, small, lightweight, and can be used easily. (See Patent Document 2)

JP 2003-255989 A (paragraphs 0014 to 0285, FIGS. 1 to 25) JP 2006-39626 A (paragraphs 0018 to 0049, FIGS. 1 to 7)

  However, it may be difficult to discriminate by a conventional spectrum analysis method that extracts a small number of features from a vibration spectrum. For example, rubber film vibration data measured using the above system is difficult to diagnose with a conventional spectrum analysis method due to the influence of strong noise due to gas turbulence and the nonlinearity of vibration measurement methods. In addition, a change in vibration level due to a change in gas flow rate has an effect on measurement data rather than a change in vibration caused by the deterioration of the rubber film. Furthermore, the difference in frequency spectrum between normal and degraded regulators is subtle. As a result, it is difficult for conventional methods such as the half-width method and linear discrimination to detect the deterioration of the rubber film at an early stage and obtain a diagnostic performance that is useful for monitoring safety prevention of the pressure regulator.

  On the other hand, the discrimination method based on the kernel method projects the given high-dimensional data to a higher-dimensional space and constitutes a linear discriminator in the space. The kernel method has been successful in various fields such as images, documents, speech classification, data mining, and bioinformatics. Therefore, rather than extracting a small number of features from the vibration spectrum, the SVM learns the shape of the entire spectrum, so that it is possible to extract subtle differences in the frequency spectrum that have been overlooked in the past, and to diagnose deterioration of the rubber film, etc. It is thought that it may be useful.

  The present invention provides a new group discriminating apparatus and group discriminating method not by extracting a small number of feature quantities from a power spectrum but by learning the shape of the whole spectrum by SVM and finding a difference from the whole spectrum. With the goal. Another object of the present invention is to provide a new group discriminating apparatus and group discriminating method that can detect deterioration of a rubber film at an early stage and are useful for monitoring safety prevention of a pressure regulator.

  In order to solve the above problem, the SVM is made to learn the shape of the entire spectrum instead of extracting a small number of features from the vibration spectrum. Also, pay attention to the KL kernel designed based on the Kullback-Leibler (KL) information amount as the kernel function. The amount of KL information is an index that is frequently used to represent the distance between two probability distributions. In the present invention, this is used as a kernel for measuring the similarity between spectra, and is applied to a group discrimination device and a group discrimination method. Is. In group discrimination, calculation time is shortened by using a calculation technique called kernel trick.

  As shown in FIG. 1, for example, the group discriminating apparatus according to claim 1 has a spectrum estimating means 2 for estimating a discrete power spectrum from the spatiotemporal sequence data using the spatiotemporal sequence data as input data, and normalizing the discrete power spectrum. Normalization means 3 for generating a normalized histogram, histogram storage means 4 capable of storing a feature vector in the normalized histogram in association with a group, and linearly or nonlinearly mapping the feature vector into a high-dimensional feature space Features of learning space-time series data having a support vector machine 6 forming a discrimination hyperplane for classifying feature vectors in a feature space using a kernel function and stored in the histogram storage means 4 using the support vector machine 6 The vector and the known group belonging to the feature vector Using the discriminant hyperplane generating means 5 for forming the hyperplane and the discriminant hyperplane formed by the discriminant hyperplane generating means 5, it is determined which group the group to which the group belongs is not yet determined And a discriminating means 7 for discriminating.

  Here, the space-time series data includes time-series data and space-series data. The time series data includes data having frequency components such as sound, light, and electric signals, and the space series data includes data having spatial frequency components such as images and patterns. The input data includes both learning data and undetermined data. In addition, the discrete power spectrum includes those obtained by variable transformation such as logarithmic transformation and affine transformation. Further, as a spectrum estimation means, a statistical spectrum estimation method such as an AR method (autoregressive method) or a periodogram method can be used. Further, the degree of discreteness of the frequency values of the discrete power spectrum may be that normally used in the periodogram method, and may be finer. Further, since the histogram storage means can store the information in association with the group, it may be stored without the association. With such a configuration, it is possible to discriminate a new group by allowing the SVM to learn the shape of the whole spectrum and finding a difference from the whole spectrum, without being limited to a certain feature amount of the extracted spectrum. Moreover, if this is applied to the determination of the deterioration of the target product, it is possible to provide a new group determination device that can find the difference at an early stage and is useful for safety prevention monitoring.

  The invention according to claim 2 is the group discriminating device according to claim 1, wherein the kernel function is any one of a KL kernel function, a chi-square kernel function, a Gaussian kernel function, a positive defined kernel function, and a polynomial kernel function. is there. When configured in this way, especially in the KL kernel function, the ability to detect minute differences between spectra such as spectral peak attenuation rate, half-width, symmetry, and shape of the spectrum between spectral peaks, compared to other kernels, The probability of discrimination is high because it is considered that the KL kernel is good. The other kernel functions listed here also have a higher discrimination probability than the conventional method.

  The invention described in claim 3 is the group discriminating apparatus according to claim 1 or 2, wherein the spectrum estimation means 2 uses either the AR method (autoregressive method) or the moving average periodogram method. The discrete power spectrum is estimated from the spatiotemporal series data. With this configuration, the probability of determination is high.

  According to a fourth aspect of the present invention, in the group discriminating apparatus according to any one of the first to third aspects, the support vector machine 6 forms a discriminant hyperplane a plurality of times, thereby forming a spatiotemporal space. Classify series data into 3 or more groups. With this configuration, for example, products can be classified into a plurality of normal, intermediate, and abnormal groups, and intermediate products can be monitored as products requiring attention. In addition, when there are a plurality of types of abnormal causes, they can be classified by cause.

  In addition, as shown in FIG. 4, for example, the group discrimination method according to claim 5 uses the learning space-time sequence data and the known belonging group related to the feature vector as input data. A first spectrum estimation step (S003) for estimating one discrete power spectrum, a first normalization step (S004) for normalizing the first discrete power spectrum to form a first normalized histogram, A first histogram storing step (S005) for storing a first feature vector in one normalized histogram in association with a group, and a kernel function for linearly mapping or nonlinearly mapping the feature vector into a high-dimensional feature space. A first histogram using a support vector machine forming a discriminant hyperplane for classifying feature vectors in the feature space; A discriminant hyperplane generating step (S006) for forming a discriminant hyperplane with the feature vector stored in the program memory step and the known group belonging to the feature vector, and the measurement target spatio-temporal series data whose group is undetermined as input data As a second spectrum estimation step (S013) for estimating the second discrete power spectrum from the measurement target spatiotemporal data, the second discrete power spectrum is normalized to form a second normalized histogram. Normalization step (S014), second histogram storage step (S015) for storing the second feature vector in the second normalized histogram, and second feature vector stored in the second histogram storage step Using the discriminant hyperplane formed in the discriminant hyperplane generation process, And a determination step of determining whether belong to any group (S016).

  Here, in the discrimination hyperplane generation step (S006) and the discrimination step (S016), calculation is efficiently performed using a calculation technique called kernel trick. With such a configuration, it is possible to discriminate a new group by allowing the SVM to learn the shape of the whole spectrum and finding a difference from the whole spectrum, without being limited to a certain feature amount of the extracted spectrum. If this is applied to the determination of the deterioration of the target product, a difference can be detected at an early stage, and a new group determination method useful for safety prevention monitoring can be provided.

  According to the present invention, a new group discriminating apparatus and a group discriminating method are provided by allowing SVM to learn the shape of the entire spectrum instead of extracting a small number of features from the power spectrum and finding a difference from the entire spectrum. it can. Moreover, if this is applied to the determination of the deterioration of the rubber film, it is possible to provide a novel group determination device and group determination method that can detect the deterioration at an early stage and are useful for monitoring safety prevention of the pressure regulator.

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

[First Embodiment]
FIG. 1 shows an example of the configuration of the group discrimination apparatus in the first embodiment. In FIG. 1, the group discriminating apparatus 1 can be constituted by a personal computer, and receives data from an analysis calculation unit 21 that performs analysis and calculation, a storage unit 22 that has a memory and stores, an external measurement system 8, a sensor, and the like. A data input unit 23 for inputting data or a user inputting data from a keyboard or the like, a display unit 24 having a display or the like for display, an output unit 25 having a printer or the like for performing printing or the like, and the entire group discrimination device 1 And a control unit 26 that performs the above control and exerts the function as a group discrimination device. The analysis calculation unit 21 includes a spectrum estimation unit 2, a normalization unit 3, a discrimination hyperplane generation unit 5, and a discrimination unit 7, and the storage unit 22 includes a histogram storage unit 4. In the present embodiment, the data input unit 23 inputs data obtained by measuring the vibration of the diaphragm (rubber film) of the high-pressure gas pressure regulator with the rubber film vibration measurement system as the measurement system 8.

  The spectrum estimation means 2 estimates a discrete power spectrum from the spatiotemporal sequence data as input data. Here, the power spectrum is estimated from the time series data detected by the microphone 40 as a sensor. The normalizing means 3 generates a normalized histogram from the discrete power spectrum. The histogram storage means 4 can store the feature vectors in the normalized histogram in association with the groups. The discrimination hyperplane generation means 5 has a support vector machine 6. This support vector machine 6 forms a discriminant hyperplane for classifying feature vectors in the feature space using a kernel function that linearly maps or nonlinearly maps the feature vectors into a high-dimensional feature space. A discrimination hyperplane is formed by the feature vectors of the learning time-space series data stored in the histogram storage means 4 and the known group belonging to the feature vectors, and is classified into a plurality of groups. The discrimination means 7 uses the discrimination hyperplane formed by the discrimination hyperplane generation means 5 to discriminate which group the belonging group belongs to the undetermined space-time series data. In this embodiment, the number of groups is 2 for simplicity.

  FIG. 2 is a sectional view showing a configuration example of the high-pressure gas pressure regulator 30 as a measurement target product in the present embodiment. There are various types of high-pressure gas pressure regulators, but the principle and basic mechanism are the same. Here, an example of a single-stage pressure regulator will be described. For example, the pressure in the high-pressure gas container is 1.56 MPa as the maximum allowable value and 0.07 MPa as the minimum allowable value, and the pressure regulator 30 uses these high-pressure gases as household gas combustors such as water heaters and stoves. The pressure is adjusted to 2.8 ± 0.5 kPa, which can be used in the above. The mechanism adjusts the opening degree of the rubber valve 35 by balancing the restoring force by the differential pressure between the air chamber 32 and the decompression chamber 33 on the atmospheric pressure side separated by the diaphragm (rubber film) 31 and the spring 34. It is. When the pressure in the decompression chamber 33 becomes higher than the target value, the diaphragm 31 is pushed up, and the displacement narrows the gap between the rubber valve 35 and the nozzle 37 via the link 36 to reduce the gas flow rate. On the other hand, when the pressure in the decompression chamber 33 becomes lower than the target value, the diaphragm 31 is lowered, the interval between the rubber valve 35 and the nozzle 37 is widened, and the gas flow rate is increased. The pressure regulator is configured as such a feedback system. Reference numeral 38 denotes a vent for maintaining communication between the air chamber 32 and the atmosphere, 39A denotes an inlet to the pressure regulator for high-pressure gas, and 39B denotes an outlet for the decompressed gas.

  There are various causes and failure modes of high-pressure gas pressure regulators. For example, in a cold region, water condensed in the air chamber 32 of the pressure regulator 30 may freeze, hindering the movement of the diaphragm 31 and causing a malfunction of the feedback system. Further, for example, the high pressure gas is reliquefied in the decompression chamber 33, disturbing the feedback system, or accelerating the deterioration of the diaphragm 31 due to the erosion of the high pressure gas. However, these failures can be detected by human hearing. In the present embodiment, a diagnostic method for early detection of abnormal vibration associated with deterioration of the diaphragm 31 that could not be detected by a conventional inspection method will be described.

  The rubber material in the pressure regulator 30 deteriorates due to environmental / operational stress such as temperature change and vibration. There are two degradation processes of rubber, one is hardening due to sublimation stressed by temperature change of the material that constitutes rubber, and the other is softening due to breakage of rubber polymer due to mechanical stress such as vibration, Both lead to a reduction in the loss factor that represents the anti-vibration properties of the rubber. Diaphragm 31 serves not only to seal high-pressure gas but also as a damper of a pressure adjustment feedback system, and rubber deterioration reduces this damping. This reduction in damping causes instability in the feedback system and induces vibrations. This sign is minor in the early stages of deterioration and cannot be detected by human hearing.

  FIG. 3 shows a configuration example of a rubber film vibration measuring system 8 as a measuring system in the present embodiment. Here, the rubber film vibration measurement system 8 includes a microphone 40, a semi-hermetic container 41, a band pass filter 42, a notch filter 43, and an amplifier 44. Since the pressure regulator 30 cannot be disassembled so that it can be easily used in the field, a method of detecting a pressure fluctuation accompanying a change in the air flow from the vent 38 provided in the air chamber 32 is adopted. A semi-sealed container 41 that covers the entire air chamber 32 is placed on the pressure regulator 30, and the pressure in the container that changes due to the air flow is detected by an inexpensive and small low-frequency broadband microphone 40 that is installed in the container. The reason for not sealing the container is to maintain communication between the air chamber 32 and the atmosphere. The waveform of the oscillating air component is distorted due to the non-linearity of the pressure regulator 30. This signal is passed through a band-pass filter 42 of 5 Hz to 500 Hz, for example, and commercial power supply noise (for example, 50 Hz) is removed through the notch filter 43. Then, the signal is amplified by the amplifier 44 and input to the group discrimination device 1. Reference numeral 45 denotes a rubber hose serving as a gas flow path. This constitutes a portable measuring means that is inexpensive, lightweight, and can be used easily.

  Next, a group discrimination method using the group discrimination apparatus in the present embodiment will be described. In the present embodiment, the groups are two groups, a normal group and an abnormal group, and the input data is time series data.

  FIG. 4 shows an example of a processing flow of the group discrimination method in the present embodiment. The processing flow is roughly divided into a learning process on the left side and a discrimination process on the right side. The learning process uses the spatiotemporal data for learning whose group is known to make the group discriminator learn the correspondence between the data and the group, and generates a discriminant hyperplane that groups them so that the learned correspondence is established. It is a process to do. The discrimination step is a step of discriminating a group to which the measurement target spatio-temporal series data whose group is undetermined is determined using the discrimination hyperplane generated by the group discrimination device.

  First, a learning process or a discrimination process is selected (selection process: S001). If it is a learning process, the learning space-time series data with which the belonging group is known is input in association with the group (learning data acquisition process: S002. In the present embodiment, four unused adjusters (hereinafter referred to as "adjustment"). A normal product, belonging to a normal group), a regulator that has passed the expiration date of 3 units for more than 3 years, and a 30-day heating test using the rubber accelerating aging test method stipulated in the JIS standard. The rubber film vibration measurement system 8 uses the adjusted rubber film for the diaphragm 31 (which belongs to an abnormal group called a deteriorated product) and a total of eight pressure regulators 30 to measure the data. Is input to the data input unit 23 of the group discriminating apparatus 1. At this time, each pressure regulator is operated four times at a flow rate of 200 L / h, 400 L / h, 600 L / h, 800 L / h, and 1000 L / h. The measurement data obtained at the sampling time 1 ms. In other words, to obtain a total of 160 data including 80 data 80 data and degradation products of normal products. Acquired data is stored in association with the group in the storage unit 22.

Next, the spectrum estimation unit 2 estimates the first discrete power spectrum from the learning spatio-temporal sequence data acquired in the learning data acquisition step (first spectrum estimation step: S003), and the normalization unit 3 The first discrete power spectrum is normalized to form a first normalized histogram (first normalization step: S004). Six methods were compared to estimate the discrete power spectrum and form a normalized histogram. The feature vector obtained by the first estimation method uses the first to thirty-order AR coefficients estimated by the Yule-Walker method (shown as ARc in Table 1), and each coefficient is expressed as x _i = [X _i (1),..., X _i (30)] ^T (T represents a transposed vector). The AR coefficient is not a spectrum, but is used for comparison because it is frequently used for time series classification. The feature vector obtained by the second estimation method uses a periodogram estimated by the discrete Fourier transform and normalized (indicated as PG in Table 1), and x _i = [x every 1 Hz from 1 Hz to 500 Hz. _{i (1), ...,} was constructed as x i ^{(500)] T.} The feature vector obtained by the third estimation method was estimated by the AR method and constituted by using a normalized spectrum (shown as AR in Table 1). The AR order here was chosen to be optimal in the meaning of the AIC (Akaike Information Criterion) (for calculation methods, Koji Akaike, Toichi Nakagawa, “Statistical Analysis and Control of Dynamic Systems”, Science Co., Ltd. Issued by the company). The structure of the feature vector is the same as that of the periodogram. The feature vectors obtained by the 4th to 6th estimation methods were constructed using periodograms normalized after smoothing by moving average of 5 points, 30 points and 60 points (in Table 1, MA1 to 3). Is displayed).

  Normalization is for handling all data with a unified standard. For example, since the difference in gas flow rate has a great influence on the power level of the spectrum, in order to eliminate the influence of the amplitude change due to the flow rate, the spectrum normalized with the total power is used.

  FIG. 5 shows normalized power spectra calculated based on vibrations measured from a normal product and a deteriorated product. FIG. 5A shows a frequency spectrum of a normal product, and FIG. 5B shows a frequency spectrum of a deteriorated product. The effective frequency range is 5 to 500 Hz. In the figure, the spectrum of 80 data for each of the normal product and the deteriorated product is superimposed and displayed. Moreover, these spectra were estimated by the AR method of the optimal order with AIC. Since these spectra have multiple peaks, it is difficult to investigate which peaks have a large impact on the diagnosis. Also, the difference in spectrum between normal and degraded products is difficult to understand at a glance. For this reason, it is not easy to apply a half width method for extracting and investigating a small number of feature amounts from the frequency peak of the conventional mode of interest.

FIG. 6 schematically shows an example of feature vectors obtained from the normalized histogram. The horizontal axis represents frequency, and the vertical axis represents the normalized power spectrum. Here, x _i (k) is the k-th frequency component of a certain discrete normalized spectrum i, and the feature vector constituted by these elements is x _i = | x _i (1),... X _i (k) ... _x i (M) ^| becomes ^T. x _j (k) is the k th frequency component of a discrete normalized spectral j, by their component including the feature vector _{x j = | x j (1} ), ... x j (k) ... x j (M) | ^T. Further, Δt in FIG. 5 is the sampling time, and M is the number of elements of the feature vector. When the spectrums of the feature vector x _i and the feature vector x _j are similar, the similarity is high, and when they are not similar, the similarity is low. That is, normal products and deteriorated products have similar spectra, and normal products and deteriorated products are expected to differ in any part of the spectrum.

  Next, the histogram storage means 4 stores the first feature vector in the first normalized histogram in association with the group (first histogram storage step: S005). In the present embodiment, feature vectors obtained from the first normalized histogram for each product are stored separately for a normal group and an abnormal group for a total of 160 data including 80 data for normal products and 80 data for degraded products. did.

  Next, the discriminant hyperplane generating means 5 uses the support vector machine 6 to form a discriminant hyperplane by using the feature vector stored in the first histogram storing step and the known group belonging to the feature vector (discriminant hyperplane). Hyperplane generation step: S006). Here, the support vector machine 6 is a learning machine that forms a discrimination hyperplane that classifies feature vectors in the feature space using a kernel function that linearly maps or nonlinearly maps the feature vectors into a high-dimensional feature space. The calculation necessary for the actual discrimination hyperplane formation is performed efficiently using a calculation technique called kernel trick.

  Here, the support vector machine (SVM) and the kernel function will be briefly described. The SVM is a linear discriminating learning machine (mainly implemented in software in a computer) for binary class discrimination. In general, if a given data set is linearly separable, there are infinite number of discriminating hyperplanes, but SVM has a maximum margin (the distance between learning data closest to the discriminating hyperplane and the hyperplane). A plane can be uniquely determined. (Refer to "Statistics of Pattern Recognition and Learning, 2 Theory and Practice of Kernel Method", published by Koji Tsuda, Iwanami Shoten Co., Ltd., published on April 11, 2003)

Now, N I / O data pairs
think of. In this case, if the discriminative hyperplane is w and the constant is β,
It can be expressed as The above equation can be formulated as equation (2) as an optimization problem in the dual space.

Here, α _i is a Lagrange multiplier, C is a normalization parameter, and <•> is an inner product. C is a parameter that controls the trade-off between the complexity of the discrimination hyperplane and the SVM error in the learning data. Here, in order to extend the SVM to a nonlinear discriminator, a nonlinear mapping Φ: X → Z from the input vector space X to the feature space Z is introduced, and the similarity between the inputs x _i and x _j within the space is < Φ (x _i ) · Φ (x _j )>. This inner product as a function of the input space
The function K (x _i , x _j ) can be expressed as a kernel function. Here, by selecting a function that does not require the calculation of the inner product in the high-dimensional superspace as the kernel function, the calculation amount of the similarity can be greatly reduced. This is a calculation technique called kernel trick. Also, formula (2)
By changing the above, the SVM becomes a nonlinear discriminator. In other words, introducing a kernel function is equivalent to constructing a linear discriminant hyperplane that takes a maximum margin in a high-dimensional space projected by the nonlinear function Φ. Therefore, the discriminating ability of the nonlinear SVM largely depends on the selection of the kernel function.

Next, the Kullback-Leibler kernel (KL kernel) will be briefly described. The KL information amount is used as an index for measuring the proximity between two probability distributions, and the distance between the random variable x and the two probability distributions p (x) and q (x) is calculated.
Define as This amount takes a value of 0 when the two distributions are the same, and it is considered that the larger the value, the farther the two distributions are. Here, since the kernel function needs to be a symmetric function, the following symmetrized KL information amount SD is considered.
If a and b are constants, the KL kernel is
Is defined as

The KL kernel has been proposed as a kernel that measures the similarity between two probability distributions. A probability model in which a mixed Gaussian model (GMM) is applied to an image component decomposed by discrete cosine transform in image discrimination is obtained, and the probability distribution and statistics are input to the KL kernel. However, since the frequency spectrum is handled here, not the stochastic model, this operation is not necessarily performed. Therefore, in order to correct the KL kernel to a format suitable for this problem, it is redefined as follows.

Here, x _i (k) is the k-th frequency component of a certain discrete normalized spectrum i as shown in FIG.

  By using the KL kernel, the support vector machine 6 can form a discriminative hyperplane for nonlinearly mapping input data to a high-dimensional feature space and classifying the input data in the feature space. Accordingly, the discriminant hyperplane generating means 5 uses the support vector machine 6 to discriminate between the feature vector generated from the learning space-time series data stored in the histogram storage means 4 and the known affiliation group related to the feature vector. A plane can be formed and classified into a normal group and an abnormal group. In addition to the KL kernel, a discrimination hyperplane can be similarly created using a Gaussian, Laplacian, sub-linear, and chi-square kernel, and products can be classified into normal groups and abnormal groups. However, the reason why the KL kernel has a higher discriminating ability than other kernels is thought to be because it has the ability to detect minute differences between spectra such as the attenuation rate, half-width, and symmetry of the spectrum peak. It is done.

  Next, it is determined whether or not all the learning space-time series data has been processed (S007). If there is unprocessed data, the process returns to the selection step (S001), and input and processing are repeated for the remaining data. If there is no unprocessed data, the process is terminated.

  If the selection step (S001) is a discrimination step, the measurement target spatio-temporal series data whose group is undetermined is input (measurement target data acquisition step: S012). In this step, the new pressure regulator 30 is measured by the rubber film vibration measuring system 8 and the data is input to the data input unit 23 of the group discriminating apparatus 1. Next, the spectrum estimation unit 2 estimates the second discrete power spectrum from the measurement target space-time series data (second spectrum estimation step: S013). Next, the normalizing means 3 normalizes the second discrete power spectrum to form a second normalized histogram (second normalization step: S014). Next, the second feature vector in the second normalized histogram is stored in the histogram storage means 4 (second histogram storage step; S015). These steps S013 to S015 are the same as steps S003 to S005, respectively, except that no association with groups is performed.

  Next, the discriminating means 7 uses the second feature vector stored in the second histogram storing step (S015) and the discriminating hyperplane formed in the classification step (S006) to measure the spatiotemporal data to be measured. Then, it is determined which group it belongs to (discriminating step: S016). In this embodiment, the product belongs to the normal group depending on whether the feature vector generated from the time series data of the vibration measured for the normal or abnormal product is determined on the discrimination hyperplane in the feature space. Or belonging to an abnormal group.

  Next, it is determined whether or not all the measurement target space-time series data has been processed (S007). If there is unprocessed data, the process returns to the selection step (S001), and input and processing are repeated for the remaining data. If there is no unprocessed data, the process is terminated.

  Next, an experiment for evaluating the discrimination ability by changing the combination of the spectrum estimation means and the kernel function will be described. The group discriminating apparatus according to the present embodiment only needs to be configured to be able to use any one of the spectrum estimation means and any one of the kernel functions. Here, all the spectrum estimation means (six types) are used. And all kernel functions (8 types) are configured to be usable.

First, extracting features from the total 160 data sets of vibration data 80 sets obtained from the vibration data 80 sets from normal regulator degradation regulator, to constitute a feature vector x _i.
In this experiment, six types of spectrum estimation means (AR unit, periodogram, spectrum estimated by AR method, periodogram smoothed by 5-point, 30-point, and 60-point moving average) and 8 types of kernel functions (1 , 2 and 3 order polynomials, Gaussian, Laplacian, sub-linear, chi-square, KL kernel), the discrimination ability of each SVM was examined.
Forty data sets (80 data sets in total) taken at random from the normal and deterioration adjuster data sets were trained by SVM as learning data sets, and trials for discriminating the remaining 80 data sets were repeated 100 times.

Table 1 shows the average misclassification rate. Table 1 also shows the smallest misclassification rate in the diagnosis by the half width method focusing on peaks near 70 Hz, 120 Hz, and 180 Hz. Here, with respect to the coefficients of Gaussian, Laplacian, sub-linear, χ2 and KL kernel, b = 0 is fixed and a = “0.01, 0.02, 0.03, 0.04, 0.05, 0 .06, 0.07, 0.08, 0.09, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 , 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 ", the average misclassification rate is the highest. It shows what was small. Symbols ARc, PG, AR, MA1, and MA2MA3 in the table are respectively smoothed by spectrum estimation means, AR coefficients, periodograms, spectra estimated by the AR method, 5-point, 30-point, and 60-point moving averages used for discrimination. P1, P2, P3, GA, SL, LA, χ ² , KL are first-order, second-order, third-order polynomials, Gaussian, Laplacian, sub-linear, χ-square, KL kernel, HP Means the half-width method. “×” is a combination that cannot be calculated by the spectrum estimation means and the kernel function. As a result of the experiment, the combination of the spectrum estimated by the AR method and the KL kernel showed the highest discrimination performance. A high discrimination performance with an accuracy rate of 92% or higher was achieved by using an inexpensive sensor and SVM.

Table 2 shows the results of discriminating experiments by estimating a probability distribution in which a Gaussian mixture model (GMM) is applied to the spectrum and using the probability distribution as an input to the KL kernel. GM3, GM4, and GM5 in the table mean GMMs using 3, 4, and 5 normal distributions, respectively. The EM algorithm (expected maximum algorithm) was used for GMM estimation. Also in this experiment, the SVM combining the spectrum estimated by the AR method and the KL kernel shows a higher discrimination ability than the SVM having the probability distribution fitted with the GMM as an input.

  The KL kernel has shown a high discrimination ability for this spectrum discrimination problem because the ability to detect minute differences between spectra such as the attenuation rate, half-value width and symmetry of the spectrum peak is higher than that of other kernels. This is probably because the kernel is good. In addition, regarding the feature vector, a spectrum smoothed moderately by the AR method or 30-point moving average smoothing shows a high discrimination rate. This is probably because the smoothness of the spectrum is insufficient in the 5-point moving average smoothing, and the spectrum is too smooth in the 60-point moving average smoothing, and the characteristics of the individual spectra are erased. Similarly, since the probability density by GMM is a weighted sum of normal distribution, it cannot cope with a minute change in spectrum. Spectrum estimation by the AR method is practically convenient because an optimum spectrum can be automatically determined by AIC.

[Second Embodiment]
In the first embodiment, an example in which a KL kernel is used as a kernel function has been described. In the second embodiment, an example in which another kernel such as a Gaussian kernel is used will be described. In the Gaussian kernel, the similarity between input data is measured by taking an exponential function by multiplying the square of the difference between each feature vector by minus. Therefore, when the similarity of each feature vector is high, the reaction is relatively insensitive to a slight change in the feature vector, but when the similarity between the feature vectors is low, the reaction is relatively sensitive to a slight change in the feature vector. In the case of the KL kernel, the ability to detect minute differences between spectra such as the attenuation rate, half width, and symmetry of the power spectrum peak is considered to be superior to other kernels. Thus, it is considered suitable for detecting a difference between dissimilar patterns.

In the Laplacian kernel, the similarity between input data is measured by multiplying the absolute value of the difference between each feature vector by minus and taking an exponential function. Therefore, compared with the Gaussian kernel, when the similarity of each feature vector is high, the attenuation of the similarity between the input data is large, but when the similarity of each feature vector is low, the attenuation of the similarity between the input data is small. Therefore, it is considered easier to detect the difference between similar patterns than the Gaussian kernel.
The sub linear kernel is measured by taking the exponential function by multiplying the similarity between input data by minus the half power of the difference between each feature vector. Therefore, it can be considered that the features of the Laplacian kernel are more prominent than those of the Laplacian kernel. Therefore, it is considered easier to detect the difference between similar patterns than the Gaussian kernel.
The d-th order polynomial kernel is formed by superposing d-th order polynomials in the discrimination hyperplane.
Further, the chi-square kernel is obtained by measuring the similarity between input data by taking the exponential function by multiplying the square of the difference between the feature vectors by scaling the sum of the feature vectors and adding the minus.
Kernels other than these KL kernels are also used for comparing differences in the whole spectrum in the group discrimination method according to the present embodiment. Compared to the conventional detection method in which only a certain feature amount is focused, other differences are also available. Can detect anomalies.

[Third Embodiment]
In the first embodiment, an example in which time series data is handled has been described. In the third embodiment, a case in which spatial series data is handled will be described. That is, a spatial frequency spectrum is used instead of the time frequency spectrum in the first embodiment. For example, measuring unevenness of the surface shape of a product using a laser positioning type, etc., generating a normalized histogram for the measured spatial data, and forming a discrimination hyperplane with a support vector machine using a kernel function By this, it is possible to classify products in which defects such as scratches have occurred on the surface shape. Further, the present invention is not limited to one-dimensional data, but can be applied to two-dimensional or more spatial data. For example, the uniformity of coating application is determined by the change in the amount of reflected light. For the two-dimensional image data, the spatial frequency spectrum can be estimated by a technique such as a wavelet transform or a two-dimensional Fourier transform. A two-dimensional image can also be handled in a one-dimensional space by tracing on the scanning line. Other configurations and processes are the same as those in the first embodiment.

[Fourth Embodiment]
Although the case where the number of groups is two groups has been described in the first embodiment, the case where the number of groups is classified into three or more groups will be described in the present embodiment. The support vector machine essentially discriminates two groups, but the case of three or more groups can be dealt with by combining a plurality of support vector machines. There are two main methods. The first method is based on one-to-other (other) classification. First, one arbitrary group is extracted and classified into a group in which the group and others are collected. Next, another arbitrary one group is extracted from the group in which other groups are collected, and is classified into a group in which the group and other groups are grouped. This is a method for performing discrimination. For example, first, data is divided into “A group and other BCD groups”, and a discrimination hyperplane is formed for the data of A vs. BCD. Similarly, “B vs. ACD”, “C vs. ABD”, and “D vs. ABC” form four discrimination hyperplanes (group number n), and discriminate target data is discriminated by each hyperplane. When the data is classified into any single group, the discrimination target data is classified into the group. In other words, in “A vs. BCD”, if it is classified into group A, it is classified into group A, and if it is classified into BCD group, it is discriminated again by the other three discriminating hyperplanes. This is continued until it is classified into a single group. (See O. Chapel, P. Haffner and V. Vanik, “SVM for Histogram-Based Image Classification”, IEEE Transactions on Neural Networks, 1999).

  The second method is based on one-to-one classification, and has a majority method and an elimination method. In the majority method, when the number of groups is n, the number of combinations of any two groups is determined as n (n-1) / 2 times, and the majority number, that is, the number of groups in which the point enters is added, and the most number of times is obtained. This is a method for determining that the user belongs to many groups. For example, “A vs. B”, “A vs. C”, “A vs. D”, “B vs. C”, “B vs. D”, “C vs. D” and 6 (number of groups = n (n−1)) A discrimination hyperplane of / 2) is prepared. Here, in the method by majority vote, the discrimination target data is discriminated by all the six hyperplanes, and the majority vote is taken as a result. In this method, the determination is completed after 6 times = n (n−1) / 2 times. In the elimination method, any two groups are discriminated, and the possibility of being classified into a group not classified in the discrimination is eliminated, and the possibility of being classified is any two other than the erased group. Judgment is made by one group. This process is repeated n-1 times, and finally classified into groups that have not been erased. For example, when “A vs. B” is determined and classified as B, the possibility of A is discarded because there is no possibility of being classified as A, and then “B vs. C”, “B” other than A group Either “D” or “C vs. D” is performed. Here, it is assumed that “B vs. C” is performed and classified into C. Here, the possibility of B disappears, and the last remaining “C vs. D” is determined. Here, the discrimination target data is classified into the finally discriminated result. In this method, the determination is completed after three times = n−1 times. (See J. C. Platt, N. Cristianini and J. S-Taylor, “Large margin tags for multi-class classification”, Advances in Neural Information Processing, p.

  By these methods, for example, it is possible to classify the pressure regulators into “normal”, “intermediate”, and “abnormal”, and the causes of deterioration such as deterioration due to hardening and softening of the diaphragm, for example. When there are two modes, it can be classified into “normal”, “deterioration due to hardening”, and “deterioration due to softening”.

  The present invention can also be realized as a program for causing a computer to execute the group discrimination method described in the above embodiment. The program may be stored and used in a built-in memory of the control unit 26 of the group discriminating apparatus 1 or may be stored and used in a storage device inside or outside the system, or downloaded from the Internet to the control unit 26 and used. Also good. Moreover, it is realizable also as a recording medium which recorded the said program.

  Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and it is obvious that various modifications can be made to the embodiments.

  For example, in the above-described embodiment, the example in which the object of application is a high-pressure gas regulator has been described. However, the present invention is also applicable to products having other vibration mechanisms or rotation mechanisms, such as water supply valves and pipes, motors, and audio equipment. Applicable. The present invention can also be applied to data having frequency components such as sound, light, and electric signals, and data having spatial frequency components such as images and patterns. Also, it is generally applicable to other spectrum classification problems without limitation. In addition, the group discriminating apparatus in the above embodiment only needs to be configured to be able to use any one spectrum estimation unit and any one kernel function. However, a plurality of spectrum estimation units and a plurality of kernels may be used. The function may be configured to be usable. Also, FFT (Fast Fourier Transform), DCT (Discrete Cosine Transform) or DFT (Discrete Fourier Transform) can be used as the spectrum estimation means. Further, the number of samples of normal products and deteriorated products, the number of feature vectors, the number of moving averages, and the like can be variously changed.

  The present invention can be used for group discrimination of spatiotemporal data. As an application, for example, it can be used for early abnormality detection of products.

It is a figure which shows the structural example of the group discrimination | determination apparatus in 1st Embodiment. It is sectional drawing which shows the structural example of a high pressure gas pressure regulator. It is a figure which shows the structural example of a rubber film vibration measuring system. It is a figure which shows the example of a processing flow of the group discrimination | determination method in 1st Embodiment. It is a figure which shows the normalized frequency spectrum of the vibration measured from the normal product and the deteriorated product. It is a figure which shows the example of a feature vector typically.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Group discrimination | determination apparatus 2 Spectrum estimation means 3 Normalization means 4 Histogram storage means 5 Discrimination hyperplane generation means 6 Support vector machine 7 Discrimination means 8 Measurement system 21 Analysis calculation part 22 Storage part 23 Data input part 24 Display part 25 Output part 26 Control unit 30 High pressure gas pressure regulator 31 Diaphragm 32 Air chamber 33 Decompression chamber 34 Spring 35 Rubber valve 36 Link 37 Nozzle 38 Vent 39A Inlet 39B Outlet 40 Microphone 41 Semi-sealed container 42 Band pass filter 43 Notch filter 44 Amplifier 45 Rubber hose

Claims (5)

Spectrum estimation means for estimating a discrete power spectrum from the spatiotemporal sequence data using the spatiotemporal sequence data as input data;
Normalizing means for normalizing the discrete power spectrum to form a normalized histogram;
Histogram storing means capable of storing feature vectors in the normalized histogram in association with groups;
A support vector machine that forms a discriminant hyperplane that classifies feature vectors in the feature space using a kernel function that linearly or non-linearly maps feature vectors into a high-dimensional feature space, and uses the support vector machine to Discriminant hyperplane generating means for forming the discriminant hyperplane with the feature vector of the learning spatio-temporal series data stored in the histogram storage means and the known group belonging to the feature vector;
Using a discrimination hyperplane formed by the discrimination hyperplane generating means, and a discrimination means for discriminating to which group the group to which the group belongs has not been determined belongs.
Group discriminator. The kernel function is any one of a KL kernel function, a chi-square kernel function, a Gaussian kernel function, a positive definitive kernel function, and a polynomial kernel function;
The group discrimination device according to claim 1. The spectrum estimation means estimates the discrete power spectrum from the spatio-temporal sequence data using either an AR method (autoregressive method) or a moving average periodogram method;
The group discriminating apparatus according to claim 1 or 2. The support vector machine classifies the spatiotemporal data into three or more groups by forming a discriminant hyperplane multiple times;
The group discrimination device according to any one of claims 1 to 3. A first spectrum estimation step of estimating a first discrete power spectrum from the learning spatiotemporal sequence data using the learning spatiotemporal sequence data and a known group belonging to the feature vector as input data;
A first normalization step of normalizing the first discrete power spectrum to form a first normalization histogram;
A first histogram storage step of storing a first feature vector in the first normalized histogram in association with the belonging group;
In the first histogram storage step, using a support vector machine that forms a discriminant hyperplane that classifies feature vectors in the feature space using a kernel function that linearly or nonlinearly maps feature vectors into a high-dimensional feature space. A discriminant hyperplane generating step of forming the discriminant hyperplane with a stored feature vector and a known group associated with the feature vector;
A second spectrum estimation step of estimating a second discrete power spectrum from the measurement target spatiotemporal sequence data using as input the measurement target spatiotemporal sequence data whose group to be determined;
A second normalization step of normalizing the second discrete power spectrum to form a second normalization histogram;
A second histogram storage step of storing a second feature vector in the second normalized histogram;
Using the second feature vector stored in the second histogram storage step and the discriminant hyperplane formed in the discriminant hyperplane generation step, which group the measurement target space-time series data belongs to A discriminating step for discriminating;
Group identification method.