Feature extraction device, feature extraction method, and program

Abstract

The present disclosure is a feature extraction device which has: an input unit for inputting time-series data which represents the system state at each point in time of a system configured from one or more devices and comprises data which has a number of dimensions equal to the number of devices constituting the system × the number of items which expresses the state of the devices; a time window generation unit for generating converted data by converting the time-series data at each point in time from the data which has the number of dimensions equal to device number × item number to data which has a number of dimensions equal to device number × item number × time window length; a feature extraction unit for generating periodgram data which expresses intensity distribution per frequency component by converting the converted data at each point in time from the data which has the number of dimensions equal to device number × item number × time window length to data which has a number of dimensions equal to device number × item number × frequency component number; and an output unit for outputting the periodgram data at each point in time.

Description

[Technical Field]

[0001] The present disclosure relates to a technique for extracting features in change-point detection. [Background technology]

[0002] There is a known technology for detecting changes in the system status of a system consisting of one or more devices using time-series data that represents the system status at each point in time. Here, "system status" refers to the operating status of the system, which is expressed by quantitative variables such as "number of accesses" and "number of users."

[0003] The technique described in Non-Patent Document 1 is known as a technique for detecting change points in time-series data to which no correct answer label is assigned regarding the position where the change point occurs.

[0004] The method of Non-Patent Document 1 is an extension of the technology for detecting change points through clustering. Because it is a clustering-based method, the target time series is not subject to constraints such as stationarity constraints or independent and identical distribution constraints. Furthermore, the method of Non-Patent Document 1 can be said to be a method that introduces the concept of a time axis, in that it clusters a time window at each point in time of time series data, then tracks the clusters assigned to each point in time along the time axis to extract transition patterns. Furthermore, the method of Non-Patent Document 1 sets a past period and a current period that are sufficiently longer than the time window, and calculates a change point score by comparing the distribution of cluster transition patterns between the two periods. This means that the method calculates change point scores for interval data with a certain time width, rather than snapshot data for each point in time, making it a technique that can detect change points, including changes in time series patterns.

[0005] Specifically, the change point detection device proposed in Non-Patent Document 1 has an input unit that inputs time series data that represents the system state at each point in time of a system composed of one or more devices, and that is composed of data with the dimension of the number of devices that compose the system x the number of items that represent the states of the devices; a time window generation unit that generates converted data by converting the time series data at each point in time from data with the dimension of the number of devices x the number of items to data with the dimension of the number of devices x the number of items x time window length; and a function (composed of a clustering unit, a cluster transition sequence creation unit, a cluster transition tensor calculation unit, and a change point score calculation unit) that detects a change point when the change point score of the system state calculated based on the converted data at each point in time exceeds a preset threshold.

[0006] Here, we consider applying Non-Patent Document 1 to time series data that exhibits periodic fluctuations. Periodic fluctuations refer to fluctuations in time zones, days of the week, seasonal fluctuations, etc., and are one of the characteristics observed in time series data in a wide range of fields, including communications. As described above, the technology in Non-Patent Document 1 is a method for detecting change points by calculating a change point score for the system state at each time point based on data of the dimensions of the number of devices x the number of items x the time window length at each time point that constitutes the time series data, and proposes passing the raw time window data directly to the detection unit. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Takahashi, T. Takeshita, "Proposal of a change detection technique using cluster transition tensor," IEICE Technical Report, CQ2021-75, pp.49-54, Nov. 2021. Summary of the Invention [Problem to be solved by the invention]

[0008] However, when detecting change points from time-series data showing periodic fluctuations, passing the raw time window data directly to the detection unit without extracting features from it leads to detecting change points without capturing the larger trends in the periodic fluctuations. The detection unit first clusters the data that make up the time-series data at each point in time, with the dimensions of the number of devices x the number of items x the time window length. However, clustering raw time window data in this way makes it impossible to capture the larger trends by ignoring the minute fluctuations in the periodic fluctuations, and results in time window data that actually show the same periodic fluctuations being classified into different clusters. This creates the problem of making it difficult to improve the accuracy of detecting change points in the system state.

[0009] The present invention has been made in view of the above points, and aims to improve the accuracy of change-point detection when applying the change-point detection device proposed in Non-Patent Document 1 to time-series data showing periodic fluctuations. [Means for solving the problem]

[0010] In order to achieve the above object, the invention of claim 1 is a feature extraction device comprising: an input unit that inputs time-series data representing a system state at each time point of a system composed of one or more devices, the time-series data being composed of data with dimensions of the number of devices constituting the system × the number of items representing the states of the devices; a time window generation unit that generates converted data by converting the time-series data at each time point from data with dimensions of the number of devices × the number of items to data with dimensions of the number of devices × the number of items × time window length; a feature extraction unit that generates periodogram data indicating an intensity distribution for each frequency component by converting the converted data at each time point from data with dimensions of the number of devices × the number of items × the number of frequency components; and an output unit that outputs the periodogram data at each time point. [Effects of the Invention]

[0011] As described above, according to the present invention, when the change-point detection device proposed in Non-Patent Document 1 is applied to time-series data showing periodic fluctuations, the effect of being able to improve the accuracy of change-point detection can be achieved. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram illustrating an example of a functional configuration of a change-point detection device that is a premise of the present embodiment. [Figure 2] 10 is a flowchart showing a change point detection process. [Figure 3] FIG. 1 is a diagram illustrating an example of a functional configuration of a feature extraction device according to an embodiment of the present invention. [Figure 4] 10 is a flowchart illustrating an example of a feature extraction process according to the present embodiment. [Figure 5] FIG. 1 is a diagram illustrating an example of a hardware configuration of a change-point detection device and a feature extraction device. DETAILED DESCRIPTION OF THE INVENTION

[0013] A change-point detection device that is the premise of this embodiment First, before describing the feature extraction device 20 of this embodiment, the change-point detection device 10, which is the premise of this embodiment, will be described with reference to FIGS. 1, 2, and 5. FIG.

[0014] Here, we will explain a change point detection device 10 that can detect the time point at which a change occurs in the system state using time series data that represents the system state at each point in time of a system (S) consisting of one or more devices. Here, the "system state" refers to the operating state of the system expressed by quantitative variables such as "number of accesses" and "number of users."

[0015] [Functional configuration] First, the functional configuration of the change-point detection device 10 will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example of the functional configuration of the change-point detection device.

[0016] 1, the change-point detection device 10 includes an input unit 11, a time window generation unit 12, a period setting unit 13, a clustering unit 14, a cluster transition sequence creation unit 15, a cluster transition tensor calculation unit 16, a change-point score calculation unit 17, a detection unit 18, and an output unit 19. Note that the "device" in the "number of devices" and "device status" shown below refers to the device that constitutes the system that is the target of change-point detection by the change-point detection device 10.

[0017] The input unit 11 inputs time series data representing the system state at each point in time of a system (S) consisting of one or more devices, and which is composed of data of dimensions (number of devices x number of items representing the device states) that make up the system (S).

[0018] The time window generation unit 12 divides the time series data input by the input unit 11 into fixed-length time windows, converts the data at each point in time from (number of devices × number of items) dimensional data to (number of devices × number of items × time window length) dimensional data, generates converted data, and performs intermediate output.

[0019] The period setting unit 13 extracts time series data of a predetermined past period and a present period from the (number of devices x number of items x time window length) dimensional time series data generated by the time window generation unit 12, and performs intermediate output.

[0020] The clustering unit 14 classifies the state of the (number of devices x number of items x time window length) dimensional data at each point in time that constitutes the time series data for the past period and the current period extracted by the period setting unit 13 using a clustering method, and produces an intermediate output.

[0021] The cluster transition sequence creation unit 15 tracks the clusters assigned by the clustering unit 14 to the (number of devices × number of items × time window length) dimensional data at each point in time in the past period and the current period, along the time axis, and creates a sequence of cluster transitions between different clusters for each of the past period and the current period, and at the same time assigns the duration of stay in that cluster to each cluster that makes up this cluster transition sequence, and produces an intermediate output.

[0022] The cluster transition tensor calculation unit 16 extracts cluster transitions of a predetermined fixed length from the cluster transition sequence created by the cluster transition sequence creation unit 15, calculates the occurrence probability of each cluster transition pattern in the past period and the current period, calculates cluster transition tensors for each of the past period and the current period, with the cluster transition length (length of cluster transition) as the rank (i.e., dimension), has unique values ​​of all clusters that appeared in the past period and the current period as indices for each dimension, and has the occurrence probability of the cluster transition pattern as its value, and performs intermediate output.

[0023] The change point score calculation unit 17 calculates the distance between the cluster transition tensor in the past period and the cluster transition tensor in the current period as the degree of change from the past period to the current period based on the cluster transition tensors for the past period and the current period calculated by the cluster transition tensor calculation unit 16, and provides an intermediate output.

[0024] The detection unit 18 detects a change point when the change point score calculated by the change point score calculation unit 17 exceeds a preset threshold. That is, the detection unit 18 detects a change point when the change point score of the system state calculated based on the data (converted data) at each time point exceeds a preset threshold.

[0025] The output unit 19 outputs the change points detected by the detection unit 18 .

[0026] [Change point detection process] Next, the change-point detection process (procedure) will be described with reference to Fig. 2. Fig. 2 is a flowchart showing an example of the change-point detection process.

[0027] Hereafter, we will assume that the number of devices that make up the system (S) is M, the number of data items that represent the system state at each point in time is K, and the number of observation points in time series data is N, and that the time series data is composed of N pieces of M×K dimensional data.

[0028] Each element of the M×K dimensional data at each point in time is K observed values ​​that represent the state of M devices at that point in time. Specifically, if the M×K dimensional data at a certain point in time is [x1, , xK, xK+1, , x2K, , x(M-1)K+1, , xMK], for example, x(m-1)K+1, , xmK for m=1, , M are the K observed values ​​of the mth device at that point in time.

[0029] Step S11: First, the input unit 11 inputs time series data consisting of N pieces of M×K (number of devices × number of items) dimensional data. That is, if the M×K dimensional data at time point n is Xn, the input unit 11 inputs time series data {X1, , XN}.

[0030] Step S12: Next, the time window generation unit 12 divides the time series data input in step S11 by a time window of fixed length W, thereby converting the data at each time point from M×K (number of devices × number of items) dimensional data to M×K×W (number of devices × number of items × time window length) dimensional data, generating converted data and performing intermediate output. Specifically, the M×K×W dimensional vector Yn=(Xn-(W-1),Xn-(W-2), ,Xn) consisting of the M×K dimensional data Xn-(W-1), Xn-(W-2), ,Xn for time points n-(W-1), n-(W-2), ,n, respectively, is defined as the M×K×W dimensional data for time point n. Note that if the original M×K dimensional data Xn is observed for time points n=1, ,N, the converted M×K×W dimensional data Yn will be obtained for time points n=W, ,N.

[0031] Step S13: Next, the period setting unit 13 extracts time series data of a preset past period and a preset current period from the M×K×W (number of devices × number of items × time window length) dimensional time series data generated in step S12. Specifically, if the past period is [s1, e1] and the current period is [s2, e2], data {Ys1, , Ye1} of the past period and data {Ys2, , Ye2} of the current period are extracted from the M×K×W dimensional data Yn at time point n=W, , N.

[0032] Step S14: Next, the clustering unit 14 obtains a cluster sequence corresponding to the time series data by classifying (e1-s1+e2-s2+2) pieces of M×K×W (number of devices × number of items × time window length) dimensional data that constitute the time series data of the past period of length (e1-s1+1) and the current period of length (e2-s2+1) extracted in step S13 using a clustering method. Specifically, when the cluster to which the M×K×W dimensional data Yn at time n belongs is defined as Cn, the clustering unit 14 obtains a cluster sequence {Cs1,...,Ce1} from the time series data of the past period {Ys1,...,Ye1} and a cluster sequence {Cs2,...,Ce2} from the time series data of the current period {Ys2,...,Ye2}. Clustering is a process of classifying (e1-s1+e2-s2+2) pieces of M×K×W dimensional data into the same cluster based on their mutual distance. A cluster sequence is obtained by arranging the clusters assigned to each M×K×W dimensional data in chronological order. The clustering method may be a hierarchical method (e.g., shortest distance method, farthest distance method, group average method, Ward's method, etc.), or a non-hierarchical method (e.g., K-Means method, etc.).

[0033] Step S15: Next, the cluster transition sequence creation unit 15 tracks the clusters assigned in step S14 for the M×K×W (number of devices × number of items × time window length) dimensional data at each point in time of the past period [s1, e1] and the current period [s2, e2] along the time axis, creates a sequence of cluster transitions between different clusters for each of the past period and the current period, and assigns the duration of stay in that cluster to each cluster that makes up this cluster transition sequence. To be more specific, let us take the example of the cluster sequence {Cs1,···,Ce1} obtained from the time series data {Ys1,···,Ye1} for the past period [s1,e1]. If the time point at which a cluster transition between different clusters occurred in the interval [s1,e1] is denoted as τi (i=1,2,···,I) (where τ1=s1), and the cluster after the transition at time point τi is denoted as c(τi), then by arranging these in chronological order, we can obtain a cluster transition sequence of length I: c(τ1) → c(τ2) →··· → c(τI). Furthermore, by assigning the duration of stay d(τi) = τi+1 - τi (where τI+1 = e1) to each cluster c(τi) that constitutes this cluster transition sequence, the cluster transition sequence with duration of stay c(τ1)|d(τ1) → c(τ2)|d(τ2) → ··· → c(τI)|d(τI) is obtained.

[0034] Step S16: Next, the cluster transition tensor calculation unit 16 extracts cluster transitions of a predetermined fixed length L from the cluster transition sequence created in step S15, calculates the occurrence probability of each cluster transition pattern in the past period and the current period, and calculates cluster transition tensors for each of the past period and the current period, with the cluster transition length L as the rank (dimension), unique values ​​of all clusters that appeared in the past period and the current period as indices for each dimension, and the occurrence probability of the cluster transition pattern as values. Specifically, taking the cluster transition sequence c(τ1) → c(τ2) → ··· → c(τI) of length I obtained from the time series data {Ys1, ···, Ye1} of the past period [s1, e1] as an example, (I-(L-1)) cluster transitions of length L (where L≦I) can be extracted from this cluster transition sequence, and are expressed as c(τi-(L-1)) → c(τi-(L-2)) → ··· → c(τi) (i=L, ···, I). The cluster transition tensor calculation unit 16 calculates the occurrence probability of these (I-(L-1)) cluster transitions for each pattern, and calculates an L-dimensional cluster transition tensor based on this. Here, the occurrence probability of a cluster transition pattern is the value obtained by dividing the occurrence frequency of the cluster transition pattern by the total occurrence frequency of all cluster transition patterns.

[0035] The frequency of occurrence of a cluster transition pattern may be a value weighted by the duration of the cluster transition pattern. Below, for simplicity's sake, the method for storing the occurrence probability of a cluster transition pattern in a tensor will be explained using an example where L = 2 and the unique values ​​of all clusters that appeared throughout the past and current periods are α, β, and γ. In this case, the cluster transition tensor is two-dimensional, with the index of each dimension taking three values: α, β, and γ. The cluster transition tensor can be represented as a 3x3 array. If the occurrence probability of the cluster transition pattern α→β is 0.1, the occurrence probability of 0.1 is stored in the array element whose index on the first axis (first element of the cluster transition pattern) takes the value α and whose index on the second axis (second element of the cluster transition pattern) takes the value β.

[0036] Step S17: Next, the change point score calculation unit 17 calculates the distance between the cluster transition tensor in the past period and the cluster transition tensor in the current period as the degree of change from the past period to the current period, based on the cluster transition tensors for the past period and the current period calculated in step S16. If the elements of the cluster transition tensor D1 for the past period are d1i1,...,iL and the elements of the cluster transition tensor D2 for the current period are d2i1,...,iL, the distance between the two can be expressed, for example, by the following mean square error: (Σl=1LΣm=1M(d2i1,...,iL-d1i1,...,iL)2 / ML)1 / 2 In the inter-tensor distance below, M is the number of unique values ​​in all clusters that appeared throughout the past and current periods.

[0037] Step S18: Next, the detection unit 18 detects a change point when the change point score calculated in step S17 exceeds a preset threshold. That is, the detection unit 18 detects a change point when the change point score of the system state calculated based on the data (converted data) at each time point exceeds a preset threshold.

[0038] Step S19: Finally, the output unit 19 outputs the change point detected in step S18.

[0039] [Hardware configuration] Next, the hardware configuration of the change-point detection device 10 will be described with reference to Fig. 5. Fig. 5 is a diagram showing the hardware configuration of the change-point detection device and feature extraction device.

[0040] 3, the change-point detection device 10 includes a processor 101, a memory 102, an auxiliary storage device 103, a connection device 104, a communication device 105, and a drive device 106. The hardware components constituting the change-point detection device 10 are connected to each other via a bus 107.

[0041] The processor 101 serves as a control unit that controls the entire change-point detection device 10, and includes various arithmetic devices such as a CPU (Central Processing Unit). The processor 101 reads various programs onto the memory 102 and executes them. The processor 101 may also include a GPGPU (General-purpose computing on graphics processing units).

[0042] The memory 102 has a main storage device such as a ROM (Read Only Memory) and a RAM (Random Access Memory). The processor 101 and the memory 102 form a so-called computer, and the processor 101 executes various programs read onto the memory 102, thereby enabling the computer to realize various functions.

[0043] The auxiliary storage device 103 stores various programs and various information used when the processor 101 executes the various programs.

[0044] The connection device 104 is a connection device that connects the change-point detection device 10 to an external device (for example, the display device 110, the operation device 111).

[0045] The communication device 105 is a communication device for transmitting and receiving various types of information to and from other devices.

[0046] The drive device 106 is a device for loading the recording medium 130. The recording medium 130 here includes media that record information optically, electrically, or magnetically, such as a CD-ROM (Compact Disc Read-Only Memory), a flexible disk, a magneto-optical disk, etc. The recording medium 130 may also include semiconductor memory that records information electrically, such as a ROM (Read Only Memory), a flash memory, etc.

[0047] The various programs to be installed in the auxiliary storage device 103 are installed, for example, by setting the distributed recording medium 130 in the drive device 106 and reading the various programs recorded on the recording medium 130 by the drive device 106. Alternatively, the various programs to be installed in the auxiliary storage device 103 may be installed by being downloaded from a network via the communication device 105.

[0048] [Major benefits of the change point detection device] As described above, the change point detection device 10 can use time series data representing the system state at each point in time of a system (S) consisting of one or more devices to detect the time point at which a change occurs in the system state as a change point.

[0049] Moreover, since the change-point detection device 10 is based on a method of classifying the system state at each time point using a clustering method, it can handle time-series data including data that does not satisfy stationarity constraints or iid constraints, such as data that shows periodic fluctuations. Furthermore, the change-point detection device 10 models the periodic fluctuations of the system (S) by considering the state transitions of the system (S) over time (i.e., the transitions between the clusters to which the system state belongs at each time point and the duration of those clusters), and can detect changes including changes in time-varying patterns such as changes in periodic fluctuations.

[0050] It should be noted that a feature extraction device 20, which will be described later, has the same hardware configuration as the change-point detection device 10, and therefore its description will be omitted.

[0051] Feature extraction device of this embodiment Next, an embodiment of the present invention will be described. In this embodiment, a feature extraction device 20 will be described that, when applying the change-point detection technology used in the change-point detection device 10 to time-series data showing periodic fluctuations, extracts features (specifically, periodograms) from input data for change-point score calculation processing (data with dimensions equal to the number of devices at each time point constituting the time-series data × the number of items × the time window length) and captures the major trends in periodic fluctuations, thereby improving the accuracy of change-point detection.

[0052] [Functional configuration] First, the functional configuration of the feature extraction device 20 according to this embodiment will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of the functional configuration of the feature extraction device according to this embodiment.

[0053] As shown in FIG. 3, a feature extraction device 20 according to this embodiment includes an input unit 21, a time window generation unit 22, a feature extraction unit 23, and an output unit 24.

[0054] The input unit 21 inputs time series data that represents the system state at each point in time of a system (S) that is composed of one or more devices, and that is composed of data of dimensions (number of devices x number of items representing the device states) that make up the system (S).

[0055] The time window generation unit 22 divides the time series data input by the input unit 21 into fixed-length time windows, converts the data at each point in time from (number of devices × number of items) dimensional data to (number of devices × number of items × time window length) dimensional data, generates converted data, and performs intermediate output.

[0056] The feature extraction unit 23 generates periodogram (intensity distribution for each frequency component) data by converting the (number of devices × number of items × time window length) dimensional data at each time point generated by the time window generation unit 22 into (number of devices × number of items × number of frequency components) dimensional data, and performs intermediate output.

[0057] The output unit 24 outputs time series data composed of data of (number of devices×number of items×number of frequency components) dimensions at each time point extracted by the feature extraction unit 23, and passes it to the detection unit 18 of the change-point detection device 10.

[0058] [Change point detection process] Next, the feature extraction process (procedure) according to this embodiment will be described with reference to Fig. 4. Fig. 4 is a flowchart showing an example of the feature extraction process according to this embodiment.

[0059] Hereafter, we will assume that the number of devices that make up the system (S) is M, the number of data items that represent the system state at each point in time is K, and the number of observation points in time series data is N, and that the time series data is composed of N pieces of M×K dimensional data.

[0060] Each element of the M×K dimensional data at each point in time is K observations that represent the state of M devices at that point in time. Specifically, the M×K dimensional data at a certain point in time is expressed as [x1, ,x K ,x K+1 ,···,x 2K ,···,x (M-1)K+1 ,···,x MK ], for example, for m=1, ,M, x (m-1)K+1 ,···,x mK are the K observations of the mth device at that time.

[0061] Step S21: First, the input unit 21 inputs time series data consisting of N pieces of M×K (number of devices×number of items) dimensional data. That is, the M×K dimensional data at time point n is converted into X n Then, the input unit 21 receives the time series data {X1, . . . , X N}.

[0062] Step S22: Next, the time window generation unit 22 divides the time series data input in step S21 by a time window of fixed length W, thereby converting the data at each time point from M×K (number of devices × number of items) dimensional data to M×K×W (number of devices × number of items × time window length) dimensional data, thereby generating converted data and performing intermediate output. Specifically, the M×K dimensional data X n-(W-1) ,X n-(W-2) ,···,X n M×K×W dimensional vector Y n =(X n-(W-1) ,X n-(W-2) ,···,X n ) is the M×K×W dimensional data at time n. Note that the original M×K dimensional data X n is observed at time points n=1, , N, the transformed M×K×W dimensional data Y n is obtained for time n=W, ,N.

[0063] Step S23: Next, the feature extraction unit 23 converts the W (time window length) dimensional data for each device and item that constitutes the M×K×W (number of devices × number of items × time window length) dimensional data at each point in time of the time series data generated in step S22 into D (=[W / 2]+1) dimensional periodogram (intensity distribution for each frequency component) data with the domain [0, [W / 2]. That is, the feature extraction unit 23 converts the M×K×W (number of devices × number of items × time window length) dimensional time series data into M×K×D (number of devices × number of items × number of frequency components) dimensional time series data, where [W / 2] is the largest integer not exceeding W / 2. The conversion to periodogram data is performed on the time window data for each device and item, but fast Fourier transform processing is applied to convert the W (time window length) dimensional time window data into D (number of frequency components) dimensional periodogram data.

[0064] Step S24: Finally, the output unit 24 outputs the M×K×D (number of devices×number of items×number of frequency components) dimensional time series data extracted in step S23 and passes it to the detection unit 18 of the change-point detection device 10 described above.

[0065] [Major Effects of the Embodiments] As described above, when applying the change-point detection technology proposed in Non-Patent Document 1 to time-series data showing periodic fluctuations, the feature extraction device 20 according to this embodiment extracts features (specifically, periodograms) in advance from input data (data with dimensions equal to the number of devices at each time point constituting the time-series data × the number of items × the time window length), converts the data into data capturing the characteristics of the periodic fluctuations, and then passes the data to the detection unit 18 of the change-point detection device 10, thereby improving the accuracy of change-point detection.

[0066] 〔supplement〕 The present invention is not limited to the above-described embodiment, and may have the following configurations or processes (operations).

[0067] The change-point detection device 10 and the feature extraction device 20 can also be realized by a computer and a program, and this program can be recorded on a (non-transitory) recording medium or provided via a network such as the Internet. [Explanation of symbols]

[0068] 10. Change point detection device 11 Input section 12 Time window generation unit 13 Period setting section 14 Clustering Department 15 Cluster transition sequence creation section 16 Cluster transition tensor calculation unit 17 Change point score calculation section 18 Detection unit 19 Output section 20 Feature Extraction Device 21 Input section 22 Time window generation unit 23 Feature extraction unit 24 Output section

Claims

1. an input unit for inputting time-series data representing a system state at each point in time of a system composed of one or more devices, the time-series data being composed of data of dimensions equal to the number of devices constituting the system multiplied by the number of items representing the states of the devices; a time window generation unit that generates converted data by converting the time series data at each time point from data with a dimension of number of devices x number of items to data with a dimension of number of devices x number of items x time window length; a feature extraction unit that converts the converted data at each time point from data with dimensions of number of devices x number of items x time window length to data with dimensions of number of devices x number of items x number of frequency components, thereby generating periodogram data that indicate an intensity distribution for each frequency component; an output unit that outputs the periodogram data at each time point; A feature extraction device having the above.

2. 2. The feature extraction device according to claim 1, wherein, when converting the transformed data at each time point from data with a dimension of the number of devices × the number of items × the time window length to data with a dimension of the number of devices × the number of items × the number of frequency components, the feature extraction unit converts the transformed data at each time point into data with a dimension of the device × the frequency component for each item by applying a fast Fourier transform to the time window length dimension data for each item related to the device × the number of items.

3. an input process for inputting time-series data representing a system state at each point in time of a system composed of one or more devices, the time-series data being composed of data in the dimension of the number of devices constituting the system x the number of items representing the states of the devices; a time window generation process for generating converted data by converting the time series data at each time point from data of the dimension of the number of devices × the number of items to data of the dimension of the number of devices × the number of items × time window length; a feature extraction process for generating periodogram data indicating an intensity distribution for each frequency component by converting the converted data at each time point from data of dimensions of number of devices × number of items × time window length to data of dimensions of number of devices × number of items × number of frequency components; an output process for outputting the periodogram data at each time point; A feature extraction method implemented by a computer.

4. A program causing a computer to execute the method according to claim 3.

Images