Anomaly detecting device, anomaly detecting method, and recording medium

1. An anomaly detecting device comprising:

a memory; and

at least one processor coupled to the memory,

the processor performing operations, the operations comprising:

when first series data are input, extracting a series feature amount being a feature amount of a signal included in the first series data;

calculating a series probability distribution being a probability distribution which the series feature amount follows;

storing a reference probability distribution being a probability distribution designated as a reference for the series feature amount in the first series data;

calculating a state feature amount representing a fluctuation condition of the series probability distribution with respect to the reference probability distribution; and

detecting an anomaly of the first series data, based on a plurality of the state feature amounts calculated from the first series data,

wherein the operations further comprise:

extracting, as the series feature amount, a predetermined feature amount of a signal at a time point of each time frame defined with a predetermined time unit,

calculating a probability distribution in which the series feature amount for each of the time frames is designated as a stochastic variable,

calculating, for each of the time frames, the state feature amount representing, by a predetermined method, a distance between the series probability distribution at a time point of the time frame and the reference probability distribution related to the time point of the time frame, and

detecting the anomaly, based on the state feature amount for each of the time frames calculated from the first series data.

2. The anomaly detecting device according to claim 1 ,

wherein the operations further

comprises

designating, as second series data, data in which the state feature amount for each of the time frames calculated from the first series data is arranged in a time-series form, statistically processing the second series data, and then detecting the anomaly.

3. The anomaly detecting device according to claim 1 ,

wherein the operations further comprises:

storing a normal model being a probability distribution of the state feature amount at a normal time in the first series data, and being a probability distribution of the state feature amount specified based on a model index signifying a time within at least a particular period, and

detecting the anomaly, based on the calculated state feature amount and the normal model.

4. The anomaly detecting device according to claim 3 ,

wherein the operations further comprises

detecting the anomaly, based on a score calculated based on a probability that the state feature amount occurs, the state feature amount indicated by the normal model related to a time point when the calculated state feature amount is acquired.

5. The anomaly detecting device according to claim 1 ,

wherein the operations further comprises:

reconstructing the state feature amount for each of the time frames defined with a predetermined time unit, according to a predetermined period, aad then generating one or more state feature amount series, designating each of the state feature amount series as second series data, statistically processing the second series data, and then detecting the anomaly.

6. The anomaly detecting device according to claim 1 , wherein a plurality of pieces of the first series data each representing a state of a different place in a target environment are input,

wherein the operations further comprises

detecting, for each piece of the first series data, the anomaly, based on a plurality of the state feature amounts calculated from the piece of the first series data.

7. The anomaly detecting device according to claim 6 ,

wherein the operations further comprises

storing a probability distribution designated as a reference for the series feature amount in the predetermined first series data.

8. The anomaly detecting device according to claim 1 , wherein a plurality of pieces of the first series data each representing a state of a different place in a target environment are input,

wherein the operations further comprises

detecting, for each piece of the first series data, the anomaly, based on a plurality of the state feature amounts calculated from the piece of the first series data.

9. The anomaly detecting device according to claim 8 ,

wherein the operations further comprises:

storing a normal model being a probability distribution of the state feature amount at a normal time in each piece of the first series data, and being a probability distribution of the state feature amount specified based on a model index signifying a time within at least a particular period, and

detecting, for each piece of the first series data, the anomaly, based on the state feature amount calculated from the piece of the first series data, and the normal model related to the piece of the first series data.

10. The anomaly detecting device according to claim 8 ,

wherein the operations further comprises:

reconstructing, for each piece of the first series data, the state feature amount calculated from the piece of the first series data according to a predetermined period, and then generating a series of one or more of the state feature amounts,

for each piece of the first series data, designating, as second series data, each of the series of the state feature amounts generated from a plurality of the state feature amounts calculated from the piece of the first series data, statistically processing the second series data, and then detecting the anomaly.

11. The anomaly detecting device according to claim 1 ,

wherein the operations further comprises:

generating the reference probability distribution by using the first series data at a normal time.

12. The anomaly detecting device according to claim 1 ,

wherein the operations further comprises:

generating a normal model by using a plurality of the state feature amounts calculated from the first series data at a normal time.

13. The anomaly detecting device according to claim 1 ,

wherein the first series data are time-series acoustic signals.

14. The anomaly detecting device according to claim 1 ,

wherein the series feature amount is a feature amount expressing a frequency and/or power of sound included in time-series acoustic signals,

expression forms of the series probability distribution and the reference probability distribution are a Gaussian mixture distribution, and

the state feature amount is a Kullback-Leibler (KL) divergence, a vector in which a predetermined number of the KL divergences are arranged, a vector in which a predetermined number of square distances of mean vectors of the Gaussian distributions in a predetermined rank are arranged, or a norm of each of the vectors.

15. An anomaly detecting method comprising:

when first series data are input, extracting a series feature amount being a feature amount of a signal included in the first series data;

calculating a series probability distribution being a probability distribution which the series feature amount follows;

calculating a state feature amount representing a fluctuation condition of the series probability distribution with respect to a reference probability distribution being a probability distribution designated as a reference for the series feature amount in the first series data; and

detecting an anomaly of the first series data, based on a plurality of the state feature amounts calculated from the first series data,

the method further comprising:

extracting, as the series feature amount, a predetermined feature amount of a signal at a time point of each time frame defined with a predetermined time unit;

calculating a probability distribution in which the series feature amount for each of the time frames is designated as a stochastic variable;

calculating, for each of the time frames, the state feature amount representing, by a predetermined method, a distance between the series probability distribution at a time point of the time frame and the reference probability distribution related to a time point of the time frame; and

detecting the anomaly, based on the state feature amount for each of the time frames calculated from the first series data.

16. A non-transitory computer-readable recording medium embodying an anomaly detecting program, the anomaly detecting program causing a computer to perform a method, the method comprising:

when first series data are input, extracting a series feature amount being a feature amount of a signal included in the first series data;

calculating a series probability distribution being a probability distribution which the series feature amount follows;

calculating a state feature amount representing a fluctuation condition of the series probability distribution with respect to a reference probability

distribution being a probability distribution designated as a reference for the series feature amount in the first series data; and

detecting an anomaly of the first series data, based on a plurality of the state feature amounts calculated from the first series data,

the method further comprising:

extracting, as the series feature amount, a predetermined feature amount of a signal at a time point of each time frame defined with a predetermined time unit;

calculating a probability distribution in which the series feature amount for each of the time frames is designated as a stochastic variable;

calculating, for each of the time frames, the state feature amount representing, by a predetermined method, a distance between the series probability distribution at a time point of the time frame and the reference probability distribution related to the time point;

and

detecting the anomaly, based on the state feature amount for each of the time frames calculated from the first series data.