Feature extraction device, feature extraction method, and program

JPWO2024232016A1Pending Publication Date: 2024-11-14
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
JP2025519231
Authority / Receiving Office
JP · JP
Patent Type
Applications
Filing Date
2023-05-09
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

Existing methods, such as Fourier transform and Koopman operators, struggle to effectively extract information about changing periodicity in time-series data, as the period is not always constant over time.

Method used

A feature extraction device using a time-dependent nonlinear mapping and QR decomposition of a positive definite kernel operator, within a Reproducing Kernel Hilbert space, to estimate a linear operator and search for a phase shift function, enabling the extraction of periodic temporal changes.

Benefits of technology

This approach allows for the characterization of periodic changes in time-series data, improving upon conventional techniques by expressing frequency information as a function, thereby facilitating the extraction of difficult-to-detect periodicity changes.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

This feature extraction device comprises: an estimation unit that is configured to perform QR decomposition of an operator Ψ composed of vectors obtained by applying a positive-definite kernel k to S time-series data items {φ1 i, ..., φn i (i = 1, ..., S) generated by φt+1 = ft (φt) on the basis of nonlinear mapping ft depending on a time on a state space Φ, and estimate, on the basis of the expression (2), a linear operator K over Mk, which is defined as the expression (1) with respect to Hilbert C*-module Mk defined by a tensor product of a space A of all continuous functions on a set of times and a reproducing kernel Hilbert space Hk related to k, and which satisfies K(av) = a(Kv) with respect to an arbitrarily defined v ∈ Mk and an arbitrarily defined a ∈ A; and a search unit that is configured to search for c1, ... , cn ∈ the set of all complex numbers, which approximately satisfy the expression (3), and a function λ ∈ A, which represents a phase shift between φt and φt+1, whereby the feature extraction device can extract a feature of a temporal change in the periodicity of the time-series data.
Need to check novelty before this filing date? Find Prior Art

Description

Feature extraction device, feature extraction method, and program

[0001] The present invention relates to a feature extraction device, a feature extraction method, and a program.

[0002] For example, much of the time-series data used in anomaly detection technology has periodicity. For example, the body temperature of an organism changes on a daily cycle. The Fourier transform has long been used as a method for extracting periodicity from time-series data, but recently, an operator-theoretic method using the Koopman operator has also been proposed (Non-Patent Document 1).

[0003] Steven L. Brunton, Eurika Kaiser, and J. Nathan Kutz, "Modern Koopman Theory for Dynamical Systems", SIAM Review, Vol. 64(2), pp. 229-340, 2022Yuka Hashimoto, Isao Ishikawa, Masahiro Ikeda, Fuyuta Komura, Takeshi Katsura, and Yoshinobu Kawahara, "Reproducing kernel Hilbert C*-module and kernel "mean embeddings", JMLR, 22, 267:1-56, 2021Senka Macesic, Nelida Crnjaric-Zic, and Igor Mezic, "Koopman Operator Family Spectrum for Nonautonomous Systems", SIAM J. applied dynamical systems, Vol. 17(4), pp. 2478-2515, 2018

[0004] However, even when time series data are periodic, the period is not always constant and may change over time. In such cases, it is difficult to extract information about how the period changes over time using existing methods such as the Fourier transform or the Koopman operator.

[0005] The present invention has been made in view of the above points, and has as its object to make it possible to extract characteristics of time-varying periodicity in time-series data.

[0006] To solve the above problem, the feature extraction device uses a time-dependent nonlinear mapping f t Based on φ t+1 = f t (φ t ) are generated by S time series data {φ 1 i , …, φ n i} (i = 1, ..., S), we apply a positive definite kernel k to the operator Ψ, which consists of vectors obtained by applying the kernel k to the set of time vectors. We decompose the operator Ψ into the space A of all continuous functions on the set of time and the reproducing kernel Hilbert space H k Let M be the Hilbert C*-module determined by the tensor product of k Regarding

[0007] and for any v∈M k and for any a∈A, K(av) = a(Kv), M k The linear operator K above is

[0008] an estimation unit configured to estimate based on

[0009] c approximately satisfies the formula 1 , ..., c n ∈ the set of all complex numbers and φ t and φ t+1 and a search unit configured to search for a function λεA that represents a phase shift of

[0010] It is possible to extract the characteristics of the time-varying periodicity of time-series data.

[0011] 1 is a diagram illustrating an example of a hardware configuration of a feature extraction device 10 according to an embodiment of the present invention; FIG. 2 is a diagram illustrating an example of a functional configuration of a feature extraction device 10 according to an embodiment of the present invention; FIG. 3 is a diagram illustrating a first estimation result for λ; and FIG. 4 is a diagram illustrating a second estimation result for λ.

[0012] In this embodiment, the Hilbert C*-module is used to extend the conventional technique using the Koopman operator, and frequency information is expressed as a function (hereinafter referred to as λ). This makes it possible to extract information (hereinafter referred to as λ) that indicates the characteristics of time-varying periodicity, which was difficult to do with conventional techniques.

[0013] [Problem Setting] Let Φ be a state space, and let us say that time series data having periodicity is generated by a discrete dynamical system expressed by the following equation (1) that depends on time (time t) on Φ.

[0014] where f t : Φ → Φ is generally a nonlinear mapping and is unknown. t is an element of Φ and is data generated at time t. In other words, equation (1) expresses data φ at time t+1. t+1 is the data φ at time t t The time-series data may be, for example, data used for anomaly detection and output from a target of anomaly detection (for example, a computer system).

[0015] H k Let be a reproducing kernel Hilbert space (Non-Patent Document 2) with respect to a positive definite kernel k on Φ. Also, let T be the set of time (a subset of all real numbers, a continuous interval containing n times t), and let A = C(T) (C(T) is the space of all continuous functions on T). M k A and H k The Hilbert C*-module (Non-Patent Document 2) is defined as the tensor product of M k The above A-linear operator K (a linear operator that represents time series data) is defined as follows:

[0016] In general, an A-linear operator L is an operator L for any v∈M k and for any a∈A, L(av) = a(Lv).

[0017] v∈H k Suppose there exists a set of λ∈A, |λ(t)|=1, and satisfying the following:

[0018] At this time, v(f t (φ t )) = λ(t)v(φ t ) holds, so λ(t) is φ t and φ t+1 (λ(t) is the time t as input, and φ t and φ t+1 (This is a function that outputs the phase shift of K.) Therefore, if K can be estimated from given time series data and λ can be obtained, it becomes possible to extract information on the time change of periodicity. Non-Patent Document 3 studies the analysis of dynamical systems that change over time using the Koopman operator, but since the equivalent of the above v also depends on time, it is difficult to extract features related to periodicity.

[0019] An embodiment of the present invention will now be described with reference to the accompanying drawings. Fig. 1 is a diagram showing an example of the hardware configuration of a feature extraction device 10 according to an embodiment of the present invention. The feature extraction device 10 in Fig. 1 includes a drive device 100, an auxiliary storage device 102, a memory device 103, a processor 104, and an interface device 105, all of which are interconnected via a bus B.

[0020] The program that realizes the processing in the feature extraction device 10 is provided by a recording medium 101 such as a CD-ROM. When the recording medium 101 storing the program is set in the drive device 100, the program is installed from the recording medium 101 to the auxiliary storage device 102 via the drive device 100. However, the program does not necessarily have to be installed from the recording medium 101, but may be downloaded from another computer via a network. The auxiliary storage device 102 stores the installed program as well as necessary files, data, etc.

[0021] When an instruction to start the program is received, the memory device 103 reads and stores the program from the auxiliary storage device 102. The processor 104 is a CPU or a GPU (Graphics Processing Unit), or a CPU and a GPU, and executes functions related to the feature extraction device 10 in accordance with the program stored in the memory device 103. The interface device 105 is used as an interface for connecting to a network.

[0022] 2 is a diagram showing an example of the functional configuration of a feature extraction device 10 according to an embodiment of the present invention. In FIG. 2, the feature extraction device 10 includes an estimation unit 11 and a search unit 12. These units are realized by processing in which one or more programs installed in the feature extraction device 10 are executed by a processor 104.

[0023] Each part will be explained below.

[0024] [Estimation Unit 11] The estimation unit 11 estimates the A-linear operator K. The method for estimating K will be described below.

[0025] Consider the following form of the positive definite kernel k:

[0026] However, k ~ (the symbol with ~ added above k) is a complex-valued positive definite kernel on Φ, and 1 is a constant function. Let ψ(φ) = k(·φ). Suppose the following S pieces of time series data generated from equation (1) are given.

[0027] For some i, Ψ = [ψ(φ 1 i ), ..., ψ(φ n i ) ]. n i ) is φ n i H obtained by applying the positive definite kernel k to k The estimation unit 11 performs QR decomposition of Ψ (Non-Patent Document 2) to obtain a Q:A such that Q = ΨR. n →Mk and R:A n →A n The estimation unit 11 then estimates K as follows:

[0028] That is, the estimation unit 11 projects K into a space created from data, thereby ~ (The symbol with ~ added above K) is obtained. However, among the S time series data, there are multiple i' and j' that satisfy the following for each j, and f t (φ j i ) can be estimated using regression or other methods.

[0029] The estimation unit 11 ~ Let be the estimate of K. ~ Each component of is an element of A (a function on T).

[0030] [Searching Unit 12] The searching unit 12 1 , ..., c n For ∈C (C is the set of all complex numbers) and λ∈A, a search is made for something that approximately satisfies the following formula (3).

[0031] Note that equation (3) is a rewrite of equation (2). k = k ~ From the construction of the positive definite kernel of 1, Q[c 1 1, ..., c n 1] is found to have the following form:

[0032] For example, the search unit 12 searches for K ~ The eigenvalues ​​of (t) are found and the regression function is used as the initial value of λ. Then, the c that minimizes the difference between the right and left sides of equation (3) is calculated. 1 , ..., c n There are methods for finding (searching) ∈C and λ∈A using the gradient method.

[0033] [Numerical Experiments] Assume that the following is true.

[0034] f tData was generated from equation (1) using various initial values, and λ was estimated using the method of this embodiment. Figure 3 shows the estimation results of λ when n = 100 and γ = π / 2 + 0.5, and Figure 4 shows the estimation results of λ when n = 200 and γ = π / 2. In Figures 3 and 4, the dashed lines indicate the correct values, and the solid lines indicate the values ​​estimated from the time-series data. Figures 3 and 4 show that the more data n used for the approximation, the better the approximation.

[0035] As described above, according to this embodiment, it is possible to extract the characteristic of the time change in periodicity of time series data as λ.

[0036] Although the embodiments of the present invention have been described in detail above, the present invention is not limited to such specific embodiments, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims.

[0037] 10 Feature extraction device 11 Estimation unit 12 Search unit 100 Drive device 101 Recording medium 102 Auxiliary storage device 103 Memory device 104 Processor 105 Interface device B Bus

Claims

1. A time-dependent nonlinear mapping f on the state space Φ t Based on φ t+1 = f t (φ t ) are generated by S time series data {φ 1 i , …, φ n i } (i = 1, ..., S), we apply a positive definite kernel k to the operator Ψ, which consists of vectors obtained by QR decomposing the operator Ψ into the space A of all continuous functions on the set of times and the reproducing kernel Hilbert space H k Let M be the Hilbert C*-module determined by the tensor product of k Regarding and for any v∈M k and for any a∈A, K(av) = a(Kv). k Let the linear operator K above be An estimation unit configured to estimate based on Approximately satisfying the formula 1 , …, c n ∈ the set of all complex numbers and φ t and φ t+1 and a search unit configured to search for a function λ∈A representing a phase shift of 2. The estimation unit, for each j in the time series data, There are multiple i' and j' that satisfy f t (φ j i 2. The feature extraction device according to claim 1, wherein the feature extraction device is configured to estimate K on the assumption that K can be estimated.

3. The search unit searches for a solution of c that minimizes the difference between the right hand side and the left hand side of the equation. 1 , …, c n 3. The feature extraction device according to claim 1, further comprising:

4. A time-dependent nonlinear mapping f on the state space Φ t Based on φ t+1 = f t (φ t ) are generated by S time series data {φ 1 i , …, φ n i } (i = 1, ..., S), we apply a positive definite kernel k to the operator Ψ, which consists of vectors obtained by QR decomposing the operator Ψ into the space A of all continuous functions on the set of times and the reproducing kernel Hilbert space H k Let M be the Hilbert C*-module determined by the tensor product of k Regarding and for any v∈M k and for any a∈A, K(av) = a(Kv). k Let the linear operator K above be An estimation procedure for estimating based on Approximately satisfying the formula 1 , …, c n ∈ the set of all complex numbers and φ t and φ t+1 A feature extraction method, characterized in that a search procedure is executed by a computer to search for a function λ∈A that represents a phase shift of A.

5. A program for causing a computer to execute the feature extraction method according to claim 4.