Signal management device and signal management method

Abstract

The objective is to improve the accuracy of the generated abnormal signal data. [Solution] The signal management device 1 includes a generator that generates pseudo-data statistically similar to true data, with each of the sequences of intensity for each frequency component corresponding to the second data being treated as true data, and a discriminator that identifies whether the pseudo-data generated by the generator belongs to the class corresponding to true data or the class corresponding to pseudo-data. The device also includes a third learning unit configured to learn multiple generative models, each having an objective function set by the probability distributions of the first data and the second data determined from the ratios estimated by the first learning unit, and the second parameter estimated by the second learning unit.

Description

[Technical Field]

[0001] The present invention relates to a signal management device and a signal management method. [Background technology]

[0002] Conventionally, techniques have been known for analyzing the characteristics of time-series signals in the frequency domain and detecting anomalies contained in the signal. For example, Patent Document 1 discloses a technique for generating data similar to the frequency spectrum of an anomaly using a generative model.

[0003] However, the technology disclosed in Patent Document 1 has many elements that depend on a combination of multiple learning conditions set during the training of the generative model, and in some cases it was difficult for the generated data to adequately reflect the characteristics of the frequency spectrum of the actual abnormal signal. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Patent No. 7710633 [Overview of the project] [Problems that the invention aims to solve]

[0005] Thus, with conventional technology, improving the accuracy of the generated abnormal signal data was a challenge.

[0006] This invention was made to solve the above-mentioned problems and aims to improve the accuracy of the generated abnormal signal data. [Means for solving the problem]

[0007] To solve the above-mentioned problems, the signal management device according to the present invention includes a first learning unit configured to learn a first parameter representing the ratio between the probability distribution of first data, which is a sequence of normal intensities for each frequency component included in the frequency spectrum of an observed signal, and the probability distribution of second data, which is a sequence of intensities for each frequency component including frequency components with intensities that deviate from the normal range of intensities, and to estimate the ratio based on the learned first parameter; and a second learning unit configured to treat each of the sequences of intensities for each frequency component included in the frequency spectrum of an observed signal as an observed value, treat each observed value as a conditionally independent discrete value, and estimate a second parameter of a class-specific probability distribution model corresponding to the class corresponding to the first data and the class corresponding to the second data, respectively, based on the frequency of occurrence of each observed value. The system comprises a second learning unit configured as described above, a generator that generates pseudo-data statistically similar to the true data, using each of the sequences of intensity for each frequency component corresponding to the second data as true data, and a discriminator that identifies whether the pseudo-data generated by the generator belongs to the class corresponding to the true data or the class corresponding to the pseudo-data, and a third learning unit configured to learn a plurality of generative models, each having an objective function set with the probability distribution of the first data and the probability distribution of the second data determined from the ratio estimated by the first learning unit and the second parameter estimated by the second learning unit, and a generation unit configured to generate the pseudo-data using each of the plurality of trained generators obtained by learning by the third learning unit.

[0008] Furthermore, the signal management device according to the present invention may further include an identification unit configured to identify whether each of the pseudo-data generated by the generation unit belongs to a class corresponding to true data or a class corresponding to pseudo-data, using each of a plurality of learned classifiers obtained by learning by the third learning unit, and a storage unit configured to store each of the pseudo-data identified by the identification unit as belonging to a class corresponding to true data.

[0009] Furthermore, the signal management device according to the present invention may further include an acquisition unit configured to acquire a series of intensities for each frequency component included in the frequency spectrum of the signal to be managed, and a determination unit configured to determine that an abnormality has occurred in the signal to be managed when the series of intensities for each frequency component included in the frequency spectrum of the signal to be managed matches the pseudo-data stored in the storage unit.

[0010] Furthermore, the signal management device according to the present invention may also include a notification unit configured to provide notification indicating the occurrence of an abnormality when the determination unit determines that an abnormality has occurred in the signal to be managed.

[0011] Furthermore, in the signal management device according to the present invention, the probability distribution of the first data is the probability density function of the first data, the probability distribution of the second data is the probability density function of the second data, and the ratio may be the density ratio of the probability density function of the first data and the probability density function of the second data.

[0012] To solve the above-mentioned problems, the signal management method according to the present invention includes a first learning step of learning a first parameter that represents the ratio between the probability distribution of first data, which is a sequence of normal intensities for each frequency component included in the frequency spectrum of the observed signal, and the probability distribution of second data, which is a sequence of intensities for each frequency component including frequency components with intensities that deviate from the normal range of intensities, and estimating the ratio based on the learned first parameter; and treating each of the sequences of intensities for each frequency component included in the frequency spectrum of the observed signal as an observed value, treating each observed value as a conditionally independent discrete value, and estimating a second parameter of a class-specific probability distribution model corresponding to the class corresponding to the first data and the class corresponding to the second data, respectively, based on the frequency of occurrence of each observed value. The system comprises a second learning step, a generator that generates pseudo-data statistically similar to the true data, with each of the sequences of intensity for each frequency component corresponding to the second data being used as true data, and a discriminator that identifies whether the pseudo-data generated by the generator belongs to a class corresponding to the true data or a class corresponding to the pseudo-data, and a third learning step that learns a plurality of generative models, each having an objective function set with the probability distribution of the first data and the probability distribution of the second data determined from the ratio estimated in the first learning step, and the second parameter estimated in the second learning step, and a generation step that generates the pseudo-data using each of the plurality of trained generators obtained by learning in the third learning step.

[0013] Furthermore, the signal management method according to the present invention may further include an identification step in which each of the pseudo-data generated in the generation step belongs to either the class corresponding to true data or the class corresponding to pseudo-data, using each of the plurality of trained classifiers obtained by learning in the third learning step; and a storage step in which each of the pseudo-data identified in the identification step as belonging to the class corresponding to true data is stored in a storage unit.

[0014] Furthermore, the signal management method according to the present invention may further include an acquisition step of acquiring a series of intensities for each frequency component included in the frequency spectrum of the signal to be managed, and a determination step of determining that an abnormality has occurred in the signal to be managed when the series of intensities for each frequency component included in the frequency spectrum of the signal to be managed matches the pseudo-data stored in the storage unit.

[0015] Furthermore, the signal management method according to the present invention may also include a notification step that provides notification indicating the occurrence of an abnormality if it is determined in the determination step that an abnormality has occurred in the signal to be managed. [Effects of the Invention]

[0016] According to the present invention, the system includes a generator that generates pseudo-data statistically similar to true data, using each of the intensity sequences for each frequency component corresponding to the second data as true data, and a discriminator that identifies whether the pseudo-data generated by the generator belongs to the class corresponding to true data or the class corresponding to pseudo-data. The system learns multiple generative models, each having an objective function set by the probability distributions of the first data and the second data determined from the ratios estimated by the first learning unit, and the second parameter estimated by the second learning unit. Therefore, the accuracy of the generated abnormal signal data can be improved. [Brief explanation of the drawing]

[0017] [Figure 1] Figure 1 is a block diagram showing the configuration of a signal management system equipped with a signal management device according to an embodiment of the present invention. [Figure 2] Figure 2 is a diagram illustrating the overview of the signal management system according to this embodiment. [Figure 3] Figure 3 is a diagram illustrating the third learning unit included in the signal management device according to this embodiment. [Figure 4] Figure 4 is a diagram illustrating the third learning unit included in the signal management device according to this embodiment. [Figure 5] Figure 5 is a diagram illustrating the third learning unit included in the signal management device according to this embodiment. [Figure 6] Figure 6 is a block diagram showing an example of the hardware configuration of the signal management device according to this embodiment. [Figure 7] Figure 7 is a flowchart showing the operation of the signal management device according to this embodiment. [Figure 8] Figure 8 is a flowchart showing the operation of the signal management device according to this embodiment. [Figure 9] Figure 9 is a flowchart showing the operation of the signal management device according to this embodiment. [Figure 10] Figure 10 is a flowchart showing the operation of the signal management device according to this embodiment. [Modes for carrying out the invention]

[0018] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to Figures 1 to 10.

[0019] [Configuration of the signal management system] First, with reference to Figure 1, an overview of the signal management system comprising the signal management device 1 according to an embodiment of the present invention will be described.

[0020] The signal management system according to this embodiment comprises a signal management device 1 and a communication terminal 2. The signal management device 1 and the communication terminal 2 are connected via a network NW.

[0021] The network NW includes, for example, wired networks such as LAN, WAN or the Internet, ISDN, mobile communication networks using wireless LAN, LTE / 4G, 5G, 6G wireless communication systems, and wireless networks such as Bluetooth (registered trademark), but the scope of the present invention is not limited to these.

[0022] The communication terminal 2 can be implemented as a mobile communication terminal such as a smartphone, a tablet computer, a laptop computer, a wearable device, etc. In this embodiment, there are n units of the communication terminal 2 (where n is a positive integer of 1 or more). The communication terminal 2 includes a terminal that is compatible with a mobile communication network and has a SIM (Subscriber Identity Module), and the contract profile of the SIM includes identifier information such as the subscriber identification number (IMSI: International Mobile Subscriber Identity).

[0023] Furthermore, the communication terminal 2 includes devices that have an IP address and are configured as IoT terminals. The communication terminal 2 is equipped with a mobile communication module and various sensors, and can detect various physical quantities and measure them as electrical signals. The communication terminal 2 transmits the measured signals via the network NW to a gateway (not shown) or a signal management device 1. In this embodiment, as an example, the reception level and signal strength of signals received from a base station, which are periodically measured and recorded by the mobile communication module of the communication terminal 2, are used as the signals to be judged for abnormalities.

[0024] As shown in region 2a of Figure 1, the communication terminal 2 measures and records time-series data of signal strength ("power [dB]"). From the waveform data of signal strength shown in region 2a, it is difficult to directly detect the occurrence of anomalies in signal strength. Therefore, the signal management device 1, described later, converts the time-series data of signal strength into a spectrum in the frequency domain and performs signal anomaly detection by analyzing the frequency components. "Strength" is a concept that includes the amplitude, power, or corresponding physical quantities of each frequency component, and also includes the use of voltage values ​​as strength.

[0025] Figure 2 illustrates the first data set, which shows a frequency spectrum containing frequency components of normal intensity, and the second data set, which shows a frequency spectrum containing frequency components of intensity that deviate from the normal range. In Figure 2, the horizontal axis represents frequency and the vertical axis represents intensity. Curve a1 shows the first data set, which represents a normal frequency spectrum with the intensity of each frequency component of a signal measured by a certain communication terminal 2. Curve b1 shows the second data set, which represents a frequency spectrum of a signal measured by a different communication terminal 2, which contains frequency components of intensity that deviate from the normal range. In the interval c of the frequency axis, the second data set b1 contains an intensity peak that is not present in the first data set a1, which represents a normal frequency spectrum, and includes an abnormal intensity that deviates from the normal range. Therefore, an anomaly is occurring in the signal with the frequency spectrum of the first data set b1.

[0026] The communication terminal 2 that measured the signal containing frequency components with an intensity that deviates from the normal range indicates a hardware or configuration problem, or a problem on the receiving environment side. For example, this could be due to a failure in the communication module of communication terminal 2, a bug or misconfiguration in the measurement software, or interference from an internal noise source. Therefore, it is considered that the communication terminal 2 that measured the signal containing frequency components with an abnormal intensity is malfunctioning.

[0027] [Functional blocks of the signal management device] Next, the functional blocks of the signal management device 1 according to this embodiment will be described with reference to the block diagram in Figure 1. As shown in Figure 1, the signal management device 1 includes an acquisition unit 10, a first learning unit 11, a second learning unit 12, a third learning unit 13, a generation unit 14, an identification unit 15, an abnormal signal database (storage unit) 16, a determination unit 17, a notification unit 18, and a storage unit 19.

[0028] The acquisition unit 10 acquires a series of intensity levels for each frequency component included in the frequency spectrum of a normal signal, measured at each of the multiple communication terminals 2 via the network NW. The acquisition unit 10 collects first data, which is a series containing frequency components of normal intensity, and second data, which is a series containing frequency components of intensity that deviate from the normal intensity range.

[0029] The acquisition unit 10 collects time-series data of signals measured at each of the multiple communication terminals 2, and performs a Fourier transform on the signals to obtain a frequency spectrum. The time-series data of the signals and their frequency spectrum are associated with the identification information of the communication terminal 2 that measured the signals. The frequency spectrum of the observed signals collected by the acquisition unit 10 has a set number of frequency components, for example, M (M=1000).

[0030] The acquisition unit 10 can label the collected frequency spectrum based on rule-based or statistical thresholds, classify it into first data and second data, and collect this data. The acquisition unit 10 acquires the first data and second data as training data for the first learning unit 11 and the second learning unit 12. The acquisition unit 10 also acquires the second data as training data for the third learning unit 13. Furthermore, the acquisition unit 10 acquires the intensity of each frequency component included in the frequency spectrum of the managed signal that the judgment unit 17 will use to determine if it is abnormal.

[0031] The first learning unit 11 learns a parameter (first parameter) that represents the ratio between the probability distribution of the first data, which is a sequence of normal intensities for each frequency component included in the frequency spectrum of the observed signal, and the probability distribution of the second data, which is a sequence of intensities for each frequency component that includes frequency components with intensities that deviate from the normal range of intensities, and estimates the ratio based on the learned parameter.

[0032] More specifically, the first learning unit 11 learns a parameter representing the density ratio of the probability density functions of the first data and the second data using the series data of the intensities for each frequency component included in the frequency spectrum of the observed signal. Further, the first learning unit 11 estimates the density ratio from the learned parameter.

[0033] Here, let the set of training data including the second data be D = {x (1) , x (2) , …, x (N)}, and the set of training data including the first data be D’ = {x’ (1) , x’ (2) , …, x’ (N’)}. x indicates the intensity for each frequency component. Each observed data x (n) is M-dimensional, and x (n) = (x1 (n) , x2 (n) , …, x M (n) ). For example, when the number of frequency components is 1000 (=M), it is 1000-dimensional. Each component x i (n) indicates the intensity at the frequency component i.

[0034] Hereinafter, D is referred to as the second data, D’ as the first data, the probability density function of the second data D as p(x), and the probability density function of the first data D’ as p’(x). In the present embodiment, the probability density function p(x) of the second data D represents a probability distribution serving as a reference distribution indicating an intensity x deviating from the range of normal intensities. Further, the probability density function of the first data D’ represents a probability distribution serving as a non-reference distribution indicating a normal intensity x, and the first learning unit 11 directly models strangeness. The density ratio r(x) between the probability density function p(x) of the second data D and the probability density function p’(x) of the first data D’ is expressed by the following equation (1).

Equation

[0035]

number

[0036] The basis function ψ(x) is defined using the RBF (Radial Basis Function) kernel and expressed by the following equation (3).

number

[0037] Here, based on equation (1) above, the specific form of equation (3) where the number of basis vectors b is the number of training data N (b=N) is given by the following equation (4).

number

[0038] Equation (4) above expresses the density ratio as a linear sum of RBFs centered on all learning points. Here, we introduce the generalized Kullback-Leibler divergence, which measures the information-theoretic distance between the non-negative functions f and g shown in equation (5).

number

[0039] In density ratio estimation, f=p(x) and g=r θ Substitute p'(x) into equation (5) above, and the objective function is given by equation (6).

number

[0040] In equation (6) above, each x n , x' n’ The optimization objective function obtained by approximating the integral with an empirical distribution that sets all values ​​except θ to 0, ignoring terms that do not depend on the parameter θ, and removing constants is expressed by the following equation (7).

number

[0041] By minimizing J(θ) in equation (7) above, the density ratio r θ This is estimated. Since J(θ) is a convex function, the first learning unit 11 updates the parameter θ from the initial value to convergence using the parameter θ update formula by gradient descent shown in equation (8) below.

number

[0042] The result of specifically calculating the gradient in equation (7) above is expressed by the following equation (9).

number

[0043] The first term of equation (9) above represents the contribution from the first data, and the second term represents the contribution from the second data. Before calculating the optimal solution for parameter θ using equations (7) to (9) above, the first learning unit 11 determines an appropriate value for the bandwidth h in equation (4) above using cross-validation or an information criterion (KL divergence minimization criterion). Based on the optimal solution for parameter θ obtained by the KL density ratio estimation method, the first learning unit 11 uses equation (2) above to calculate the density ratio r for any input x. θ An estimate of this can be obtained.

[0044] The second learning unit 12 treats each of the intensity sequences for each frequency component as an observed value, and treats each observed value as a conditionally independent discrete value. Based on the frequency of occurrence of each observed value, it estimates the parameters (second parameters) of the class-specific probability distribution model corresponding to the class of the first data and the class of the second data, respectively. The second learning unit 12 sets up a class-specific probability distribution model using a multinomial distribution for the intensity data for each frequency component of the first data and the intensity data for each frequency component of the second data, estimates its parameters, and calculates the log-likelihood of the sequence of each observed value based on the class-specific probability distribution model. The second learning unit 12 uses integer discrete values ​​as observed values ​​of the intensity of each frequency component in the intensity observed for multiple frequency components.

[0045] The second learning unit 12 focuses on the fact that the simple Bayes method treats each observation as a conditionally independent discrete variable, and extends the simple Bayes method to a multinomial distribution model to set up a class-based probability distribution model. Here, as shown in equation (10), the observed value X is a vector representing the number of occurrences of M categories for the data corresponding to the intensity of each frequency component. X = x = (x1, x2, ..., x M ) ···(10)

[0046] Furthermore, the class label to which the observed value X belongs, i.e., the event Y, is defined as a binary variable by the following equation (11).

number

[0047] In equation (11) above, the class of abnormal intensity is the class that represents intensity that deviates from the normal intensity range, and the class of normal intensity is the class that represents normal intensity. In this embodiment, the second learning unit 12 defines the class of abnormal intensity as event Y=1 and the class of normal intensity as event Y=0 in order to directly learn and evaluate the characteristics of intensity that deviates from the normal intensity range, using the normal intensity of the frequency component as the reference state. When each of the conditional probabilities P(X|Y) that occur when event Y occurs are independent, the observed variable x under the given conditions of event Y i The variables are conditionally independent of each other, and the likelihood ρ(x|y) is given by the simple Bayesian estimation method in equation (12).

[0048]

number

[0049] In this embodiment, since we are dealing with a sequence of intensities for each frequency component of integer values, we are dealing with a vector of occurrences for each category x=(x1,x2,···,x) rather than a single observed value. M We consider ) as one sample and extend it to a multinomial distribution probability model while maintaining the naive assumption that the intensity data for each frequency component are conditionally independent. The occurrence probabilities θ1,···,θ for each marginal probability P(X) in which the observed value X is observed are given. M Under the constraints that the variables are independent and the sum of their probabilities of occurrence is 1, the probability model of the multinomial distribution can be expressed by the following equation (13).

[0050]

number

[0051] As shown in equation (13) above, the observed value x is the count value of the data corresponding to the intensity of each frequency component. i The sum of (x1 + x2 + ... + x M Once the values ​​are determined, the distribution becomes the product of the following equation (14), and the observed value x corresponds to the intensity of the i-th frequency component in the series. i Each with an independent probability θi xi It is possible to find this.

number

[0052] Therefore, it can be seen that the relationship is similar to that of the simple Bayesian method in (12). Here, the prior probability P(Y) is a binary problem between the class of the first data (normal intensity) (Y=0) and the class of the second data (abnormal intensity) (Y=1), and the unknown parameters of each are θ. 0 ,θ 1 Let's define D as the marginal probability (result), and the observed value of the first data (normal intensity) (Y=0) be D. 0 The observed value of the second data (abnormal intensity) (Y=1) is D 1 Let's assume that equation (14) above can be expressed as equation (15) by transforming the product form into a sum under the Naive Bayes independence assumption and decomposing the log-likelihood for the data set, the set of observed values ​​D, into class-specific forms.

[0053]

number

[0054] Here, the constraint is expressed by equation (16).

number

[0055] Furthermore, applying the Lagrangian multiplier method, the parameter θ related to the class of the first data (normal intensity) is obtained. 0 The maximum value of the log-likelihood for the i-th component is given by equation (17).

number

[0056] Parameter θ related to the class of the second data (abnormal intensity) 1 Similarly, when we find the maximum value of the log-likelihood for this as well, under the constraint of equation (16) above, the parameter θ corresponds to the class of the second data (abnormal intensity) and the class related to the first data (normal intensity), respectively. 1 θ 0 This can be expressed by the following equation (18).

number

[0057] Note that x i Since x takes an integer value, i (n) To prevent the multinomial distribution from diverging when θ is 0, smoothing can be performed by specifying +α (e.g., α=1) smoothing. In this way, the second learning unit 12 considers the sequence of observed data corresponding to the intensity of each frequency component to follow a multinomial distribution, estimates the parameters of the class-specific probability distribution model by counting the frequency of occurrence of each category, and uses the likelihood based on this model to evaluate the degree of normal intensity represented by the first data and the degree of abnormal intensity represented by the second data. The parameter θ estimated by the second learning unit 12 1 θ 0 It will be handed over to the 3rd Learning Department, 13.

[0058] The third learning unit 13 includes a generator 131 that generates pseudo-data statistically similar to the true data, using each of the observed intensity sequences corresponding to the second data as true data, and a discriminator 132 that identifies whether the pseudo-data generated by the generator 131 belongs to the class corresponding to the true data or the class corresponding to the pseudo-data. The third learning unit 13 learns multiple generative models, each having an objective function set with the probability distribution of the first data and the probability distribution of the second data determined from the density ratio estimated by the first learning unit 11, and the parameters (second parameters) estimated by the second learning unit 12.

[0059] As shown in Figure 3, the third learning unit 13 performs adversarial learning on a Generative Adversarial Network (GAN) having a generator 131 and a discriminator 132. As shown in Figure 3, the third learning unit 13 performs adversarial learning on each of the multiple Generative Adversarial Networks (GANs) provided for each observed value i (i=1,...,M) corresponding to the intensity of each frequency component. Through the learning of the third learning unit 13, a total of M trained generators 131' are constructed, each corresponding to an observed value 1 to M for the intensity of each frequency component.

[0060] Figures 4 and 5 schematically represent the neural network configuration of the GAN generator 131 and discriminator 132 used by the first learning unit 11. As shown in Figure 4, the generator 131 consists of a neural network having an input layer, a hidden layer, and an output layer. The generator 131 is a model that generates pseudodata from random noise. For example, m randomly sampled Gaussian noise vectors are input to the input nodes of the generator 131 (z1~z m ).

[0061] The generator 131 outputs an output G(z) after performing a sum-of-products operation on the input and weight parameters, followed by thresholding using an activation function. The output G(z) from the generator 131 is pseudo-data similar to the observed value corresponding to the intensity of the i-th frequency component corresponding to the second data. A CNN or ResNet can be used as the neural network that constitutes the generator 131.

[0062] The classifier 132 shown in Figure 5 consists of a neural network having an input layer, a hidden layer, and an output layer. In the example in Figure 5, the input to the training data is an observed value x corresponding to the intensity of the i-th frequency component corresponding to the second data, which is acquired by the acquisition unit 10.

[0063] The classifier 132 outputs a probability value in the range of 1 to 0 after performing a sum-of-products operation on the input and weight parameters and thresholding using an activation function. When the classifier 132 correctly identifies the training data relating to the input true data as true data, it outputs a probability value close to y=1. On the other hand, when it correctly identifies the training data relating to the input pseudo-data as pseudo-data, it outputs a probability value close to y=0. In this way, the classifier 132 is a model that distinguishes the model distribution generated by the generator 131 from the data distribution of the training data, which is the true distribution. A CNN can be used as the neural network that constitutes the classifier 132.

[0064] Figure 3 is a block diagram illustrating adversarial learning of the GAN by the third learning unit 13. The generator 131 of the GAN adopted by the first learning unit 11 is denoted as function G, and the discriminator 132 is denoted as function D. The true data is denoted as x, the predicted value output by the discriminator 132 is denoted as y, and the correct label is denoted as t. The correct label t is set to 1 for the true data and 0 for the pseudo-data generated by the generator 131. In this case, the discriminator 132 is a binary classification problem with the cross-entropy E given by equation (19) below. CE It can be expressed as follows.

[0065]

number

[0066] The first term inside the brace in equation (19) above represents t n lny n In this case, the predicted value y of the classifier 132 n However, the true correct label for the data is t n It is desirable to approach the value of =1. On the other hand, the second term inside the brace represents (1-t n )ln(1-y n In this case, the predicted value y of the classifier 132 n However, the value of the correct label (1-t) that distinguishes the pseudodata from the real data. n It is desirable for cross-entropy E to approach 0. CE This value is maximized when the predicted value matches the correct label value.

[0067] Here, the generator 131 that constitutes the GAN has parameter w G ,θ G It has a function G(w G ,θ G ) is expressed as. Also, the classifier 132 has parameter w D ,θ D It has a function D(w D ,θ D This is expressed as ). The cross-entropy E in equation (19) above CE The objective function E of a GAN comprising a generator 131 and a discriminator 132 based on the above can be expressed by the following equation (20).

number

[0068] The first term of equation (20) above represents E D(x)=1 lnD(w D ,θ D ) is the expected value that the classifier 132 identifies as true data. The second term of equation (20) above represents E D(x)=0 ln(1-D(G(w G ,θ G ),w D ,θ D )) is the expected value that the classifier 132 identifies the pseudodata generated by the generator 131 as pseudodata. Here, the expected value of equation (20) above can be expressed as equation (21) using a probability distribution.

[0069]

number

[0070] Here, for the probability density function p(x) of the second data (abnormal intensity) and the probability density function p'(x) of the first data (normal intensity) in equation (1) above, we set p(x) ≡ ρ(x|y=1) and p'(x) ≡ ρ(x|y=0) in order to make them into a probabilistic labeled classification problem for min-max learning of the GAN. The density ratio r estimated by the first learning unit 11 θ(x) is defined by the following equation (22).

number

[0071] The probability density function ρ(x|y=0), which is the conditional probability distribution of the first data (normal intensity) in equation (22) above, can be calculated using the maximum likelihood estimation method or the like from the intensity data for each frequency component of the first data, which has been acquired in advance by the acquisition unit 10. Using the calculated probability density function ρ(x|y=0) of the first data, the probability density function ρ(x|y=1), which is the conditional probability distribution of the second data (abnormal intensity), can be expressed by the following equation (23).

number

[0072] Substitute the probability density function ρ(x|y=1) of the second data and the probability density function ρ(x|y=0) of the first data in equation (23) into the objective function E in equation (21), and set the prior probability ρ(y=1) for the abnormal intensity class and the prior probability ρ(y=0) for the normal intensity class to the parameter θ corresponding to the observed value x corresponding to the intensity of the i-th frequency component corresponding to the normal intensity class and the abnormal intensity class estimated by the second learning unit 12. 1 θ 0 We use the following. Furthermore, the posterior probability ρ(y=1|x) that the observed value x corresponding to the intensity of the i-th frequency component is an abnormal intensity is given by D(w) in the optimal solution of equation (22) above. D ,θ D ) corresponds to the observed value x corresponding to the intensity of the i-th frequency component, and the posterior probability ρ(y=0|x) that it is a normal intensity is given by 1-D(G(w) in the optimal solution of equation (21) above. G ,θ G ),w D ,θ D This corresponds to the density ratio r. θ And it can be determined from the prior probability.

[0073] In the learning of the GAN, the generator 131 and the discriminator 132 are adversarially learned by min-max optimization of the objective function E. Therefore, the generator 131 is learned so as to generate pseudo-data that can deceive the discriminator 132, and the discriminator 132 is learned so as to identify the pseudo-data generated by the generator 131 as pseudo-data.

[0074] In the learning of the discriminator 132, when true data is given, the discriminator 132 outputs an output close to y = 1 to maximize the first term of the objective function E in the above equation (21). On the other hand, when pseudo-data is given, learning is performed so as to maximize the second term of the objective function E by the discriminator 132 outputting an output close to y = 0.

[0075] In the learning of the generator 131, D(G(w G ,θ G ),w D ,θ D )(D(G(z)) in FIG. 3) is minimized by outputting G(w G ,θ G )(G(z) in FIG. 3) such that it approaches 1. The third learning unit 13 uses a learning procedure that alternately updates the parameters of the generator 131 and the parameters of the discriminator 132. Details of the learning procedures of the generator 131 and the discriminator 132 by the third learning unit 13 will be described later.

[0076] When the objective function E of the GAN is optimized, the third learning unit 13 passes the learned generator 131' constructed for each observation value i corresponding to the intensity of the frequency component to the generation unit 14. Therefore, finally, M learned generators 131' are constructed.

[0077] The generation unit 14 generates pseudo-data using each of the learned generators 131' obtained by the learning by the third learning unit 13. Therefore, the generation unit 14 generates a set of M pseudo-data.

[0078] The identification unit 15 uses each of the multiple trained classifiers 132' obtained through training by the third learning unit 13 to identify whether each of the pseudo-data generated by the generation unit 14 belongs to the class corresponding to true data or the class corresponding to pseudo-data. Each trained classifier 132' outputs a probability value indicating whether each of the pseudo-data generated by the trained generator 131' is true data or pseudo-data. The identification unit 15 identifies the data as belonging to the true data (abnormal intensity) class if the probability value output by each trained classifier 132' is close to 1, and identifies it as belonging to the pseudo-data (normal intensity) class if the probability value is close to 0.

[0079] The abnormal signal database 16 stores each of the pseudo-data generated by the generation unit 14 that has been identified by the identification unit 15 as belonging to a class corresponding to true data. The abnormal signal database 16 stores pseudo-data for each communication terminal 2. In this way, the abnormal signal database 16 stores the frequency spectrum of frequency components whose intensity deviates from the normal intensity range, i.e., can be considered abnormal data, as a template for abnormality detection.

[0080] The determination unit 17 determines that an abnormality has occurred in the managed signal if the sequence of intensity of frequency components included in the frequency spectrum of the managed signal acquired by the acquisition unit 10 matches the pseudo-data stored in the abnormal signal database 16. The determination unit 17 can determine that an abnormality has occurred in the managed signal because at least a portion of the sequence of intensity of frequency components included in the frequency spectrum of the managed signal matches the pseudo-data stored in the abnormal signal database 16, which indicates that the managed signal contains frequency components of abnormal intensity.

[0081] The determination unit 17 compares the intensities of each frequency component in the pseudo-data, starting with the lowest frequency component, and can determine that an anomaly has occurred in the managed signal when frequency components with matching intensities are found. The determination unit 17 can also compare the pseudo-data stored in the anomaly signal database 16 with the intensity series of each frequency component of the managed signal and make a determination based on whether the similarity meets a predetermined threshold. The threshold is a criterion for determining the similarity between the intensity series of each frequency component of the managed signal and the pseudo-data stored in the anomaly signal database 16, and can be arbitrarily set based on, for example, the distribution of frequency component intensity series observed in the past, or the conditions for generating pseudo-data. Even when using a threshold, the comparison may be performed based on the entire intensity series of each frequency component, or based on the intensities of some of the frequency components included in the series.

[0082] The notification unit 18 issues a notification indicating the occurrence of an abnormality when the determination unit 17 determines that an abnormality has occurred. The notification unit 18 can send an alarm to an external operation center (not shown). Alternatively, the notification unit 18 may send a notification via the network NW to the communication terminal 2 where the abnormality was detected.

[0083] The memory unit 19 stores the parameters and density ratios estimated by the first learning unit 11, as well as the parameters estimated by the second learning unit 12. The memory unit 19 stores M trained generators 131'.

[0084] [Hardware configuration of the signal management device] Next, an example of a hardware configuration for realizing the signal management device 1 having the functions described above will be explained using Figure 6.

[0085] As shown in Figure 6, the signal management device 1 can be implemented, for example, by a computer equipped with a processor 102 connected via a bus 101, main memory 103, communication interface 104, auxiliary storage 105, and input / output I / O 106, and a program that controls these hardware resources. Furthermore, the signal management device 1 includes a display device 107.

[0086] The processor 102 is a circuit or device that performs arithmetic processing, and can be implemented by, for example, a general-purpose central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), etc. Alternatively, some or all of these may be combined to form the processor.

[0087] The main memory 103 is composed of, for example, volatile random access memory (RAM), and pre-stores programs for the processor 102 to perform various controls and calculations. The processor 102 and the main memory 103 work together to realize the various functions of the signal management device 1, such as the acquisition unit 10, the first learning unit 11, the second learning unit 12, the third learning unit 13, the generation unit 14, the identification unit 15, the determination unit 17, and the notification unit 18 shown in Figure 1.

[0088] The communication interface 104 is an interface circuit for networking the signal management device 1 with various external electronic devices.

[0089] The auxiliary storage device 105 consists of a read / write storage medium and a drive device for reading and writing various information such as programs and data to the storage medium. The auxiliary storage device 105 can use non-volatile storage such as a hard disk or flash memory as the storage medium.

[0090] The auxiliary storage device 105 has a program storage area for storing the signal management program. The auxiliary storage device 105 also has a program storage area for storing parameters representing the density ratio of the probability density functions of the first and second data, and a first learning program for estimating the density ratio, which is executed by the signal management device 1. The auxiliary storage device 105 also has a program storage area for storing a second learning program for estimating parameters using polynomial Bayes, which is executed by the second learning unit 12. The auxiliary storage device 105 also has a program storage area for storing a third learning program for performing adversarial learning of a generative model, which is executed by the third learning unit 13.

[0091] The auxiliary storage device 105 enables the implementation of the abnormal signal database 16 and storage unit 19 described in Figure 1. Furthermore, it may also have, for example, a backup area for backing up the aforementioned data and programs.

[0092] The I / O106 is an input / output device that accepts signals from external devices and outputs signals to external devices.

[0093] The display device 107 is composed of an organic EL display or a liquid crystal display. The display device 107 displays information related to signal management, such as a sequence of intensities for each frequency component included in the frequency spectrum of the signal being managed, and the judgment result.

[0094] [Signal management device operation] Next, the operation of the signal management device having the above-described configuration will be explained with reference to the flowcharts in Figures 7 to 10.

[0095] As shown in Figure 7, first, the acquisition unit 10 acquires, from the series of intensity for each frequency component included in the frequency spectrum of the signal recorded by the communication terminal 2, a series of intensity observed for each frequency component corresponding to the first data relating to normal intensity, and a series of intensity observed for each frequency component corresponding to the second data relating to intensity deviating from the normal intensity range (step S1).

[0096] Next, the first learning unit 11 performs first learning processing (step S2). FIG. 8 is a flowchart for explaining the first learning processing in step S2. As shown in FIG. 8, the first learning unit 11 learns a parameter θ representing a density ratio r between the probability density function of the first data and the probability density function of the second data (step S20). The first learning unit 11 updates the parameter θ by the gradient descent method or the like using the above equations (7) to (9), and obtains an optimal solution of the parameter θ. θ Then, based on the above equation (2), the first learning unit 11 estimates the density ratio r from the learned parameter θ obtained in step S20 (step S21).

[0097] Returning to FIG. 7, the second learning unit 12 executes second learning processing (step S3). FIG. 9 is a flowchart for explaining the second learning processing in step S3. As shown in FIG. 9, the second learning unit 12 sets parameters θ, θ for observation values corresponding to the intensities of respective frequency components related to the abnormal intensity class and the normal intensity class (step S31). Subsequently, the second learning unit 12 defines the log-likelihood according to the above equation (13) (step S32). In step S32, the second learning unit 12 changes the objective function for optimization from the product form (Equation (14)) to the sum form. Next, the second learning unit 12 uses the Lagrange multiplier method of the above equation (17) to estimate the parameter θ corresponding to the class of the second data (abnormal intensity) that maximizes the log-likelihood (Equation (15)) under the constraint (Equation (16)), and the parameter θ corresponding to the class of the first data (normal intensity) (Equation (18)) (step S33). The estimated parameters θ, θ are stored in the storage unit 19, and the process proceeds to step S4 in FIG. 7. θ

[0098] 1 0 <() 1 0 1 0

[0099] ​​​​​​​​Next, the third learning unit 13 executes the third learning process (step S4). In step S4, the third learning unit 13 includes a generator 131 that generates pseudo-data statistically similar to the true data, using observed intensity values ​​for each frequency component corresponding to the second data as true data, and a discriminator 132 that identifies whether the pseudo-data generated by the generator 131 belongs to the class corresponding to the true data or the class corresponding to the pseudo-data. The third learning unit 13 learns multiple generative models, each having an objective function set with the probability density function of the first data and the probability density function of the second data determined from the density ratio estimated by the first learning unit 11 in step S2, and the parameters estimated by the second learning unit 12 in step S3.

[0100] Figure 10 is a flowchart illustrating the third learning process in step S4. As shown in Figure 10, first, the third learning unit processes the density ratio r estimated in the first learning process in step S2 of Figure 7. θ The probability density function ρ(x|y=1) of the second data and the probability density function ρ(x|y=0) of the first data (equation (21) above), determined from the above, are set as the objective function E of the GAN (equation (21)) (step S50). More specifically, the third learning unit 13 calculates the probability density function ρ(x|y=0), which is the conditional probability distribution of the first data, from the first data acquired by the acquisition unit 10 in step S1, using the maximum likelihood estimation method or the like. The third learning unit 13 also determines the probability density function ρ(x|y=1), which is the conditional probability distribution of the second data, expressed by equation (23) above, from the calculated probability density function ρ(x|y=0) of the first data.

[0101] Next, the third learning unit 13 calculates the parameters θ corresponding to the abnormal intensity class and the normal intensity class, respectively, which were estimated in the second learning process of step S3 in Figure 7. 1 ,θ 0 The prior probability ρ(y=1) for abnormal intensity and the prior probability ρ(y=0) for normal intensity are set as the objective function E of the GAN (Equation (22)) (Step S51). Furthermore, in the optimal solution of the objective function E in Equation (21) above, D(w D ,θ D) becomes the posterior probability ρ(y=1|x) that the observed value x corresponding to the intensity in the frequency component is an abnormal intensity, and 1-D(G(w G ,θ G ),w D ,θ D ) is the posterior probability ρ(y=0|x) that the observed value x corresponding to the intensity of the frequency component is normal intensity. These posterior probabilities ρ(y=1|x) and ρ(y=0|x) are obtained from the estimated density ratio r θ It can also be calculated from the prior probabilities ρ(y=1) and ρ(y=0).

[0102] Next, the third learning unit 13 acquires each of the intensity sequences observed for each frequency component corresponding to the second data acquired in step S1 of Figure 7 as true data (step S52). In step S52, the observed intensity value at the frequency component i=1, which corresponds to one of the observed intensity values ​​(i=1) of the M frequency component intensity observation values ​​corresponding to the second data, is taken as true data.

[0103] Next, the third learning unit 13 inputs the true data, which is the observed intensity value of the i=1 frequency component related to the second data acquired in step S52, as training data 134 into the classifier 132, and sets the parameter w of the classifier 132 so that the true data can be distinguished from the true data (y=1). D ,θ D Learn and update (step S53). As shown in the block diagram of the third learning unit 13 in Figure 3, the true data is used as training data 134 that is input when training the classifier 132.

[0104] In step S53, the third learning unit 13 can train the classifier 132 on true data using, for example, the backpropagation method. Step S53 pre-constructs a classifier 132 that can distinguish true data from other true data.

[0105] Next, the third learning unit 13 generates Gaussian noise and provides a random vector of the generated Gaussian noise as input to the generator 131 (step S54). Subsequently, the generator 131 uses the input z and weight parameter w based on the given Gaussian noise. G ,θ G The sum-of-products operation and thresholding using an activation function are performed to generate pseudo-data G(z) (step S55).

[0106] Next, the third learning unit 13 trains the classifier 132. The training of the classifier 132 involves the parameters w of the generator 131. D ,θ D This is done with fixed parameters. First, the third learning unit 13 provides the true data acquired in step S52 as training data 134 to the classifier 132 as input. Then, the third learning unit 13 uses backpropagation or the like to maximize the objective function E in equation (22) above, and the parameters w D ,θ D Update (step S56). Note that the label for training data 134 is set to 1 (true data).

[0107] Next, the third learning unit 13 provides the pseudo-data generated by the generator 131 in step S55 as input to the discriminator 132, and processes the parameters w using backpropagation or the like so that the objective function E in equation (21) above is maximized. D ,θ D Update (step S57). That is, in steps S56 and S57, in order to maximize the objective function E in equation (21) above, the first term is D(w D ,θ D )=1 is output, and the second term is D(G(w G ,θ G ),w D ,θ D The optimization is performed so that ) = 0. Note that the training data 134 has a label of 0 (pseudodata).

[0108] The learning of the classifier 132 in steps S56 and S57 corresponds to the dashed arrows in the block diagram of the third learning unit 13 shown in Figure 3, which indicate that the classifier error is calculated in block 135 of the objective function E based on the output 133 from the classifier 132, and then the error is backpropagated to the classifier 132.

[0109] Next, the third learning unit 13 trains the generator 131. During the training of the generator 131, the parameters of the discriminator 132 are fixed. The third learning unit 13 trains the generator 131 so that when random Gaussian noise is supplied to the generator 131, pseudo-data is generated. Specifically, the first learning unit 11 uses methods such as backpropagation to minimize the objective function E in equation (21) above, thereby adjusting the parameters w G ,θ G Update (Step S58).

[0110] The learning in step S58 corresponds to the dashed arrow flow in the block diagram of the third learning unit 13 in Figure 3, which indicates that the error is backpropagated to the generator 131. In other words, step S58 corresponds to the dashed arrow flow in which the pseudo-data generated by the generator 131 in Figure 3 is input to the discriminator 132, the generator error is calculated from its output 133 in block 135 of the objective function E, and then the error is backpropagated to the generator 131.

[0111] Subsequently, the learning of the discriminator 132 and generator 131 from steps S55 to S58 is repeated until the value of the objective function E reaches a Nash equilibrium and converges (step S59: NO). On the other hand, if the value of the objective function E converges (step S59: YES), the processing from steps S53 to S59 is repeated until the generator 131 and discriminator 132 are learned using all the true data (step S60: NO). Note that the processing from steps S52 to S58 and from steps S59 to S60 can be performed in batches.

[0112] Subsequently, if the generator 131 and discriminator 132 have been trained using all the true data (step S60: YES), the third learning unit 13 stores the trained generator 131' corresponding to the observed intensity value at frequency i=1 in the storage unit 19 (step S61). Further, as shown in connector a, adversarial learning is performed on each of the generative models constructed for each of the observed intensity values ​​at frequency i=2, ..., M (step S61: NO), and after storing M trained generators 131' in the storage unit 19 (step S61: YES), the process moves to step S5 in Figure 7.

[0113] Next, the generation unit 14 generates pseudo-data corresponding to the observed intensity values ​​of each frequency component using each of the trained generators 131' constructed by the third learning unit 13 (step S5). Subsequently, the identification unit 15 identifies whether each of the pseudo-data generated in step S5 belongs to the class corresponding to true data or the class corresponding to pseudo-data (step S6). Then, the abnormal signal database 16 stores each of the pseudo-data identified in step S6 as belonging to the class of true data (step S7).

[0114] Subsequently, the acquisition unit 10 acquires from the communication terminal 2 a series of intensity values ​​observed for each frequency component included in the frequency spectrum of the managed signal measured by the communication terminal 2 (step S8). Next, the determination unit 17 determines that an abnormality has occurred in the managed signal if the series of intensity values ​​observed for each frequency component included in the frequency spectrum of the managed signal acquired in step S8 matches the pseudo-data stored in the abnormal signal database 16 in step S7 (step S9).

[0115] Subsequently, if the determination unit 17 determines in step S9 that an abnormality has occurred in the managed signal, the notification unit 18 notifies an operation center (not shown) of an alarm via the network NW (step S10).

[0116] As described above, according to the signal management device of this embodiment, the probability density function and parameters obtained by the first learning unit 11 and the second learning unit 12 are set as the objective function of the GAN, adversarial learning of the GAN is performed using each of the observed values ​​corresponding to the intensity of each frequency component corresponding to the second data as true data, and pseudo-data is generated using each of the generated trained generators 131'. Therefore, the accuracy of the generated abnormal signal data can be improved.

[0117] According to the signal management device 1 of this embodiment, a generation model is constructed for each observed value corresponding to the intensity of each of the M frequency components. The objective function for each generation model corresponding to each observed value is set to include parameters for the class of normal intensity (first data) and the class of abnormal intensity (second data) corresponding to the observed value related to the intensity of the frequency component estimated by the second learning unit 12. Therefore, the learning accuracy of the generation model can be further improved.

[0118] In the embodiment described, the first learning unit 11 was configured to learn parameters representing the ratio of the probability distributions of the first data and the second data, and the case where the density ratio between probability density functions was estimated was described. However, the probability distribution is not limited to a probability density function, but may be represented as a discrete distribution, a mixture distribution, or a model that approximates these. Furthermore, the first learning unit 11 is not limited to learning the density ratio representing the ratio of the probability distribution of the first data and the second data, but may be configured to learn one or more parameters including a weighting coefficient, likelihood ratio, score value, or evaluation index for distinguishing between the first data and the second data, derived based on the ratio.

[0119] Furthermore, in the embodiments described, the third learning unit 13 was explained using the example of a case where the generative model has the configuration of a GAN, but the generative model is not limited to this. For example, the generative model may be constructed based on any generative model that learns the distribution of the second data and generates pseudo-data, such as a VAE (Variational Autoencoder) or an Energy-Based Model (EBM). In this case as well, the discrimination process for determining whether the generated pseudo-data is statistically equivalent to the real data is not limited to the configuration included in the generative model, and may be performed by a separate classifier or evaluation model.

[0120] Although embodiments of the signal management device and signal management method of the present invention have been described above, the present invention is not limited to the embodiments described above, and various modifications that a person skilled in the art can envision are possible within the scope of the invention described in the claims. [Explanation of symbols]

[0121] 1...Signal management device, 2...Communication terminal, 10...Acquisition unit, 11...First learning unit, 12...Second learning unit, 13...Third learning unit, 14...Generation unit, 15...Identification unit, 16...Abnormal signal database, 17...Determination unit, 18...Notification unit, 19...Storage unit, 101...Bus, 102...Processor, 103...Main memory, 104...Communication interface, 105...Auxiliary memory, 106...Input / output I / O, 107...Display device, 131...Generator, 132...Identifier, 133...Output, 134...Training data, 135...Block of objective function E.

Claims

1. A first learning unit is configured to learn a first parameter that represents the ratio between the probability distribution of first data, which is a sequence of normal intensities for each frequency component included in the frequency spectrum of an observed signal, and the probability distribution of second data, which is a sequence of intensities for each frequency component including frequency components with intensities that deviate from the normal range of intensities, and to estimate the ratio based on the learned first parameter. A second learning unit is configured to take each of the series of intensity values ​​for each frequency component included in the frequency spectrum of the observed signal as an observed value, treat each observed value as a conditionally independent discrete value, and estimate the second parameter of the class-specific probability distribution model corresponding to the class corresponding to the first data and the class corresponding to the second data, respectively, based on the frequency of occurrence of each observed value. A third learning unit is configured to learn multiple generative models, each having an objective function set by the probability distributions of the first data and the second data determined from the ratios estimated by the first learning unit, and the second parameters estimated by the second learning unit. The third learning unit includes a generator that generates pseudo-data statistically similar to the true data, with each of the series of intensity for each frequency component corresponding to the second data being used as true data, and a discriminator that identifies whether the pseudo-data generated by the generator belongs to the class corresponding to the true data or the class corresponding to the pseudo-data, and each objective function set by the probability distributions of the first data and the second data determined from the ratios estimated by the first learning unit. A generation unit configured to generate the pseudo-data using each of the multiple trained generators obtained by the learning of the third learning unit, A signal management device equipped with the following features.

2. In the signal management device according to claim 1, Furthermore, an identification unit is configured to use each of the multiple trained classifiers obtained through training by the third learning unit to identify whether each of the pseudo-data generated by the generation unit belongs to the class corresponding to the true data or the class corresponding to the pseudo-data, A storage unit configured to store each of the pseudo-data identified by the identification unit as belonging to a class corresponding to true data, A signal management device equipped with the following features.

3. In the signal management device according to claim 2, Furthermore, the acquisition unit is configured to acquire a series of intensities for each frequency component included in the frequency spectrum of the signal being managed, Furthermore, a determination unit is configured to determine that an abnormality has occurred in the managed signal when the sequence of intensity of frequency components included in the frequency spectrum of the managed signal matches the pseudo-data stored in the storage unit. A signal management device characterized by comprising the following features.

4. In the signal management device according to claim 3, Furthermore, the system includes a notification unit configured to provide notification indicating the occurrence of an abnormality when the determination unit determines that an abnormality has occurred in the signal being managed. A signal management device characterized by the following features.

5. In the signal management device according to claim 1, The probability distribution of the first data is the probability density function of the first data, The probability distribution of the second data is the probability density function of the second data, The ratio is the density ratio of the probability density function of the first data to the probability density function of the second data. A signal management device characterized by the following features.

6. A first learning step involves learning a first parameter that represents the ratio between the probability distribution of first data, which is a sequence of normal intensities for each frequency component included in the frequency spectrum of the observed signal, and the probability distribution of second data, which is a sequence of intensities for each frequency component including frequency components with intensities that deviate from the normal range of intensities, and estimating the ratio based on the learned first parameter. A second learning step involves treating each of the series of intensity values ​​for each frequency component included in the frequency spectrum of the observed signal as an observed value, treating each observed value as a conditionally independent discrete value, and estimating the second parameter of the class-specific probability distribution model corresponding to the class corresponding to the first data and the class corresponding to the second data, respectively, based on the frequency of occurrence of each observed value. A third learning step involves learning multiple generative models, each having an objective function set by the probability distributions of the first data and the second data determined from the ratios estimated in the first learning step, and the second parameters estimated in the second learning step. The objective function is set by the probability distributions of the first data and the second data determined from the ratios estimated in the first learning step. A generation step in which pseudo-data is generated using each of the multiple trained generators obtained through the learning in the third learning step, A signal management method comprising the following features.

7. In the signal management method described in claim 6, Furthermore, the identification step involves using each of the multiple trained classifiers obtained through the learning in the third learning step to identify whether each of the pseudo-data generated in the generation step belongs to the class corresponding to the true data or the class corresponding to the pseudo-data. A storage step in which each of the pseudo-data identified in the identification step as belonging to a class corresponding to the true data is stored in a storage unit; A signal management method comprising the following features.

8. In the signal management method described in claim 7, Furthermore, the acquisition step involves obtaining a series of intensity levels for each frequency component included in the frequency spectrum of the signal being managed, Furthermore, a determination step is made to determine that an abnormality has occurred in the managed signal if the sequence of intensity of frequency components included in the frequency spectrum of the managed signal matches the pseudo-data stored in the storage unit. A signal management method characterized by comprising the following:

9. In the signal management method described in claim 8, Furthermore, if the determination step determines that an abnormality has occurred in the managed signal, the system includes a notification step that provides notification indicating the occurrence of the abnormality. A signal management method characterized by the following:

Images