Patent Application
20220289131
The present invention relates to an anomaly estimation device, an anomaly detection device, an anomaly detection method, and an anomaly detection program in time series data in which a feature pattern may change over time.
An example of time series data in which a feature pattern may change over time is data related to vibration acquired by a sensor mounted on a vehicle. A typical example of such an in-vehicle sensor is a dashboard camera, which has a mechanism that, in response to detection of an acceleration of more than a preset threshold value (e.g., 0.2 G), determines that the vehicle has crashed and thus detects an anomaly.
In the present invention, as typical time series data, time series data related to vibration acquired by an in-vehicle sensor is handled. However, the time series data to be handled is not limited to time series data related to vibration acquired by an in-vehicle sensor as long as it is time series data in which a feature pattern may change over time. A technique of identifying an anomaly from general time series data is called anomaly detection and change detection, for which several methods have been proposed in, for example, NPL 1. Further, NPL 2 discloses a technique related to dynamic mode analysis.
However, fora dashboard camera, it is difficult to detect a weak crash event of less than the threshold value (e.g., the side surface of the vehicle body having been rubbed against a wall, the bottom surface of the vehicle body having been rubbed against a curb, etc.) as an anomaly. Therefore, the date and time when the vehicle body may be damaged cannot be identified unless all images from the dashboard camera are checked. Therefore, for a vehicle available for multiple users (e.g., a shared car, a rental car, a company car, etc.), even if the vehicle manager finds a scratch on the vehicle body, it is necessary to check the images continuously until an image when the scratch was made is found in order to check who was driving the vehicle. In addition, it is unclear as to whether the techniques disclosed in NPL 1 and NPL 2 can be used as techniques for detecting an anomaly from time series data, such as time series data related to vehicle vibration, in which a feature pattern may change over time.
A first object of the present invention is to provide an anomaly estimation device capable of limiting the images to be checked to investigate the cause of a scratch on a vehicle available for multiple users. A second object of the present invention is to provide an anomaly detection device capable of acquiring a degree of anomaly required for anomaly detection even when a feature pattern of time series data changes over time.
An anomaly estimation device of the present invention is an anomaly estimation device for estimating the time when an anomaly occurred in a vehicle available for multiple users. The anomaly estimation device of the present invention includes a sensor, a prediction unit, an anomaly degree acquisition unit, and an anomaly determination unit. The sensor is installed in a vehicle to acquire time series data related to vibration. The prediction unit predicts, based on first section data that is data of a first section that is a partial section of the time series data, data of a second section that is a partial section of the time series data later than the first section to obtain predicted second section data. The anomaly degree acquisition unit acquires a degree of anomaly in the second section based on a difference between the predicted second section data and second section data that is actual data of the second section in the time series data. The anomaly determination unit determines whether or not an anomaly has occurred in the second section based on the acquired degree of anomaly.
An anomaly detection device of the present invention is an anomaly detection device that detects a degree of anomaly in time series data. The anomaly detection device of the present invention includes a first acquisition unit, a prediction unit, and an anomaly degree acquisition unit. The first acquisition unit acquires, from a first section that is a partial section of the time series data, a dynamic feature pattern of the first section. The prediction unit predicts data of a second section that is a partial section of the time series data later than the first section by using the feature pattern to obtain predicted second section data. The anomaly degree acquisition unit acquires a degree of anomaly based on a difference between the predicted second section data and actual data of the second section in the time series data.
According to the anomaly estimation device of the present invention, at least a range where there is no anomaly can be excluded, so that it is possible to limit images to be checked to investigate the cause of a scratch on a vehicle available for multiple users. According to the anomaly detection device of the present invention, the predicted second section data is obtained based on the dynamic feature pattern, so that it is possible to acquire the degree of anomaly required for anomaly detection even when the feature pattern of the time series data changes over time.
Embodiments of the present invention will be described below in detail. Note that components having the same function are given the same numeral, and duplicate description will be omitted.
<Candidate 1 for a Solution Different from the Present Invention>
As a solution, a large amount of observation data for normal and observation data for anomaly are prepared to generate probability distributions.
In(p(x|y=1)/p(x|y=0)),
which is a degree of anomaly. However, since the situation inside a vehicle changes from moment to moment, it is not realistic to collect a large amount of data that covers all normal patterns and all anomaly patterns. Therefore, it is difficult to estimate the normal and anomaly probability distributions. As a result, Candidate 1 for a solution cannot be used in the present invention.
<Candidate 2 for a Solution Different from the Present Invention>
Time series data x(1), . . . , x(K(M+N−1)) is divided by every M+N−1 sections to generate sections D1, D2, . . . , Dk, . . . , Dx which have elements of the time series data. Dk is a section having section data x (1+(k−1)(M+N−1)), . . . , x(k(M+N−1)). Note that Dk may be expressed as:
Dk=[x(1+(k−1)(M+N−1)) . . . , x(k(M+N−1))]T, which is a column vector having M+N−1 elements. The section data Dk is divided into N partial time series which is each a partial time series having a window length (length of the partial time series) M and a shift width of 1. The nth partial time series xk,n of the kth section Dk has M elements of time series data x(n+(k−1) (M+N−1)), . . . ,x(n+(k−1)(M+N−1)+M−1), and is expressed as:
x
k,n=[x(n+(k−1)(M+N−1)), . . . ,x(n+(k−1)(M+N−1)+M−1)]T,
which is a column vector. Note that for restrictions on expression, the column vector is expressed as above, but this expression means the same as:
The same applies to the following. Next, N elements of partial time series xk,1, . . . , xk,N of the section Dk are put together into a data matrix Xk as:
X
k=[Xk,1 . . . xk,n . . . xk,N].
Note that, although the matrix is expressed as above due to restrictions on expression, this expression means the same as:
The lines extending up and down of xk,1, xk,n, and xk,N indicate that xk,1, xk,n, and Xk,N are each a column vector.
Here, the kth section Dk is set as the current section (second section), and the section Dk-L earlier by L sections is set as a past section (first section) to be compared.
U
k-L,r=[uk-L,1,uk-L,2, . . . ,uk-L,r].
Here, uk-L,1, uk-L,2, . . . , uk-L,r are each a column vector. Similarly, a matrix of singular vectors qk,1, qk,2, . . . , qk,r extracted from the data matrix Xk-L for the second section is defined as:
Q
k,r=[qk,1,qk,2, . . . ,qk,r].
Here, qk,1, qk,2, . . . , qk,r are each a column vector. Note that r may be set to a different value between the first section and the second section.
A degree of change c(k) is:
c(k)=1−//Uk-L,rTQk,r//2.
Here, // //2 represents a matrix 2-norm, and //Uk-L,rTQk,r//2 is equal to the maximum singular value of Uk-L,rTQk,r.
In the technique of NPL 1, a feature pattern is extracted by singular value decomposition for the section of interest. This corresponds to decomposing the eigenvalues of a variance matrix of a data matrix for the section of interest. Accordingly, since only the section of interest is focused on and direction vectors are extracted such that their variance becomes large, only the static feature patterns assuming that the section of interest is stationary are successfully extracted. However, in an environment where the situation changes from moment to moment, such stationarity cannot be assumed, so that it is believed that features fail to be extracted correctly and thus the error in the degree of change becomes large, resulting in an increase in false detection.
<Anomaly Estimation Device of the Present Invention>
The first acquisition unit 110 acquires a dynamic feature pattern of the first section (past section) Dk-L, that is a partial section of the time series data x(i) (S110). The dynamic feature pattern is a feature pattern related to a periodic vibration and a damping rate of the periodic vibration. In other words, what periodic vibration causes changes over time and what damping (increase) rate the periodic vibration has are extracted. More specifically, a dynamic eigenmode m (matrix in which M-dimensional column vectors are arranged for the number of ranks) and an eigenvalue Λ=diag(λ1) for the first section are obtained by dynamic mode decomposition. Next, the dynamic mode decomposition will be described. Note that the “damping rate” described herein includes a case where it increases when the positive or negative sign is reversed.
Similar to the description of Candidate 2 with reference to
X
k-L,n=[x(n+(k−L−1)(M+N−1)), . . . ,x(n+(k−L−1)(M+N−1)+M−1)]T,
The first acquisition unit 110 generates a matrix Xk-L,1=[xk,1 . . . xk,n . . . xk,N-1] and a matrix Xk-L,2=[xk,2 . . . Xk,n . . . xk,N]. Note that the matrix Xk-L,1 and the matrix Xk-L,2 are expressed as the above for convenience of expression, but these expressions mean the same as the following:
The first acquisition unit 110 uses a low-rank approximated pseudo-inverse matrix Xk-L,1+ obtained based on the result of singular value decomposition for Xk-L,1 to obtain a linear operator A as:
A=X
k-L,2
X
k-L,1
+.
Then, the first acquisition unit 110 obtains an eigenmode Φk-L and an eigenvalue Λk-L so as to satisfy:
A=Φ
k-LΛk-LΦk-L−1
Note that the detailed procedure for obtaining the linear operator A from the singular value decomposition using a pseudo-inverse function is as follows. First, performing singular value decomposition on Xk-L,1,
a singular matrix Σk-L,1 and unitary matrices Uk-L,1, Vk-L,1 are obtained as:
X
k-L,1
=U
k-L,1Σk-L,1Vk-L,1T.
Approximating these up to the number of ranks r results in:
X
k-L,1,r
=U
k-L,1,rΣk-L,1,rVk-L,1,rT.
The low-rank approximated pseudo-inverse matrix Xk-L,1* can be obtained as:
X
k-L,1
+
=V
k-L,1,rΣk-L,1,r−1Uk-L,1,rT.
Thus, by the singular value decomposition and low-rank approximation for Xk-L,1, the linear operator A can be obtained as:
A=X
k-L,2
V
k-L,1,rΣk-L,1,r−1Uk-L,1,rT.
In this way, the low-rank approximation makes it possible to reduce the computation cost.
Further, the detailed procedure for obtaining the eigenmode Φk-L and the eigenvalue Λk-L is as follows. Again, the low-rank approximation is used to reduce costs. First,
˜A is obtained as:
˜
A=U
k-L,1,r
T
AU
k-L,1,r
=Uk-L,1,rXk-L,2Vk-L,1,rΣk-L,1,r−1.
By eigenvalue decomposition,
a matrix W in which rows of eigenvectors are arranged to satisfy
˜
AW=WΛ
and an eigenvalue Λk-L=diag(λj) are obtained. Then, the unique mode Φk-L is obtained as:
Φk-L=Xk-L,2Vk-L,1,rΣk-L,1,r−1W.
The eigenmode Φk-L and the eigenvalue Λk-L obtained in this way satisfy the following relation:
AΦ
k-L=Φk-LΛ.
The prediction unit 120 predicts, based on first section data that is data of the first section Dk-L that is a partial section of the time series data, data of the second section Dk that is a partial section of the time series data later than the first section Dk-L to obtain predicted second section data (S120). Specifically, the data of the second section is predicted using the first partial time series xk,l of the second section and the feature pattern (eigenmode Φk-L and eigenvalue Λk-L), and the predicted second section data {circumflex over ( )}x(1+(k−1)(M+N−1)), . . . , {circumflex over ( )}x(N+(k−1) (M+N−1)+M−1) is obtained. As described above, the eigenmode Φk-L and the eigenvalue Λk-L are acquired by the first acquisition unit 110. The prediction unit 120 obtains {circumflex over ( )}xk,n for n=2, . . . , N as:
{circumflex over ( )}xk,n=Φk-LΛk-Ln-1Φk-L−1xk,1.
Alternatively, the prediction unit 120 may first obtain
b=Φ
k-L
−1
x
k,1, and
obtain {circumflex over ( )}xk,n for n=2, . . . , N as:
{circumflex over ( )}xk,n=Φk-LΛk-Ln-1b.
Then, the prediction unit 120 may set the first elements of the partial time series xk,1 and the predicted partial time series {circumflex over ( )}xk,2, . . . , {circumflex over ( )}Xk,N-1 as the first to (N−1)th data {circumflex over ( )}x(1+(k−1)(M+N−1)), . . . , {circumflex over ( )}x(N−1+(k−1)(M+N−1)) of the predicted second section data, and set all elements of the predicted partial series {circumflex over ( )}xk,N as the Nth to (M+N−1)th data {circumflex over ( )}x(N+(k−1)(M+N−1)), . . . , {circumflex over ( )}x(N+(k−1)(M+N−1)+M−1) of the predicted second section data. Note that, although “{circumflex over ( )}” is described in front of “x” due to restrictions on expression, it has the same meaning as “{circumflex over ( )}” placed on the top of “x” as illustrated in
The anomaly degree acquisition unit 130 acquires a degree of anomaly in the second section based on a difference between the predicted second section data and second section data that is actual data of the second section in the time series data (S130). Specifically, the anomaly degree acquisition unit 130 may acquire, from the predicted second section data {circumflex over ( )}x(1+(k−1)(M+N−1)), . . . , {circumflex over ( )}x(N+(k−1)(M+N−1)+M−1) and the actual data x(1+(k−1)(M+N−1)), . . . , x(N+(k−1)(M+N−1)+M−1) of the second section in the time series data, a degree of anomaly c(k) as:
c(k)=(({circumflex over ( )}x(1+(k−1)(M+N−1))−x(1+(k−1)(M+N−1)))2+({circumflex over ( )}x(2+(k−1)(M+N−1))−x(2+(k−1)(M+N−1)))2+, . . . ,+({circumflex over ( )}x(t+(k−1)(M+N−1))−x(t+(k−1)(M+N−1)))2+, . . . ,+(({circumflex over ( )}x(k(M+N−1))−x(k(M+N−1)))2)1,2.
Note that the above expression is:
as having the same meaning.
The anomaly determination unit 220 determines whether or not an anomaly has occurred in the second section based on the acquired degree of anomaly (S220). For example, a threshold value for the degree of anomaly may be set, and when there is a difference exceeding the threshold value, it may be determined that an anomaly has occurred.
In the first embodiment, the application is limited to a vehicle available for multiple users such as a shared vehicle and a rental car. However, the present invention can be widely applied to time series data in which a feature pattern may change over time. In other words, the data to be acquired by the sensor does not have to be limited to data related to vibration. Accordingly, in a second embodiment, an anomaly detection device will be described that receives input of some time series data and outputs a degree of anomaly for which an anomaly is likely to be detected.
To the anomaly detection device 20, time series data x(i) is input (S211). The first acquisition unit 110 acquires, from the first section (past section) Dk-L, that is a partial section of the time series data, a dynamic feature pattern of the first section (S110). The prediction unit 120 predicts, based on first section data that is data of the first section Dk-L that is a partial section of the time series data, data of the second section Dk that is a partial section of the time series data later than the first section Dk-L to obtain predicted second section data (S120). The anomaly degree acquisition unit 130 acquires a degree of anomaly in the second section based on a difference between the predicted second section data and second section data that is actual data of the second section in the time series data (3130). Steps S110, S120, and S130 are the same as in the first embodiment.
In the method of obtaining the degree of change from a degree of overlap between the eigenmode of the past section and the eigenmode of the current section as described as Candidate 2 in the first embodiment, an eigenvalue representing a damping (increase) rate is not taken into consideration. Therefore, even if there is an eigenmode with the same frequency, in a case where a damping feature pattern is for the past section and an increase feature pattern is for the current section, they are distinguished only by comparing their eigenmodes, so that a degree of change available for anomaly detection cannot be acquired.
According to the present invention, a dynamic feature pattern of the section of interest can be estimated only from time series data, so that the accuracy of estimating a degree of change of the time series data obtained from a system in which the situation changes dynamically is improved, and thus the detection accuracy of the changed part is improved. Since the feature pattern extracted in the observation section is a signal corresponding to an eigenmode (eigenfrequency), the data can be easily interpreted. Therefore, according to the anomaly detection device of the present invention, the predicted second section data is obtained based on the dynamic feature pattern, so that it is possible to acquire the degree of anomaly required for anomaly detection even when the feature pattern of the time series data changes over time.
[Program, Storage Medium]
The various types of processing described above can be performed in a manner that a program for executing the respective steps of the above method is loaded on a storage unit 2020 of a computer illustrated in
The program describing the processing content can be stored in a computer-readable storage medium. The computer-readable storage medium may be, for example, a magnetic storage device, an optical disk, a magneto-optical storage medium, a semiconductor memory, or the like.
Further, the distribution of such a program is performed, for example, by selling, transferring, or renting a portable storage medium such as a DVD or CD-ROM in which the program is stored. In addition, the program may be stored in a storage device of a server computer, and then the program may be distributed by being transferred from the server computer to another computer via a network.
A computer that executes such a program first temporarily stores, for example, the program stored in a portable storage medium or the program transferred from a server computer in its own storage device. Then, to execute processing, the computer reads the program stored in its own storage medium and executes the processing according to the read program. Further, as another execution form of this program, a computer may read the program directly from a portable storage medium and execute processing according to the program, and also, every time the program is transferred from a server computer to this computer, processing according to the received program may be executed sequentially. Further, the above-mentioned processing may be executed by a so-called ASP (Application Service Provider) type service, which implements processing function only by executing a program in accordance with an instruction and acquiring a result without transferring the program from a server computer to this computer. Note that the program in this form includes information to be used for processing performed by a computer and equivalent to the program (data that is not a direct command to the computer but has a property for defining the processing of the computer, etc.).
Further, in this form, the present device is configured by executing a predetermined program on a computer, but at least a part of these processing contents may be realized by hardware.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2019/030717 | 8/5/2019 | WO |
