Abnormality estimation device, abnormality estimation method, and program
View Patent ↗Provided is an abnormality estimation device capable of appropriately determining normal data appearing less frequently as normal. The abnormality estimation device includes an estimation unit that estimates an anomaly degree of an acoustic signal, by using an abnormality estimation model that is optimized while using a set of normal sounds and is optimized so as to minimize a difference between an anomaly degree of a normal sound appearing more frequently and an anomaly degree of a normal sound appearing less frequently.
1. An abnormality estimation device that estimates an anomaly degree of an acoustic signal being input, comprising:
processing circuitry configured to
estimate the anomaly degree of the acoustic signal, by using an abnormality estimation model that is optimized while using a set of normal sounds and is optimized so as to minimize a difference between an anomaly degree of a normal sound appearing more frequently and an anomaly degree of a normal sound appearing less frequently, wherein the set of normal sounds comprises the normal sound appearing more frequently and the normal sound appearing less frequently, wherein
the abnormality estimation model is trained by applying a larger weight to the normal sound appearing less frequently among the set of normal sounds than to the normal sound appearing more frequently thereby reducing false positives.
2. The abnormality estimation device according to claim 1 , wherein
while U(x) expresses a uniform distribution having a constant probability density with the set of normal sounds and having no density in other regions, whereas A θ (x) expresses an abnormality estimation model to calculate a likelihood of an observation vector x being anomalous by using a parameter θ, the processing circuitry estimates the anomaly degree of the acoustic signal, by using the abnormality estimation model A θ (x) optimized so as to minimize an objective function J θ expressed as
[
Math
.
26
]
J
θ
=
∫
U
(
x
)
A
0
(
x
)
dx
.
3. The abnormality estimation device according to claim 2 , wherein the anomaly degree of the acoustic signal is estimated by:
approximating the objective function J θ as
[
Math
.
27
]
J
θ
≈
1
N
∑
n
=
1
N
1
p
(
x
n
)
A
0
(
x
n
)
where a true generative model of the observation vector x is expressed as p(x), and n=1, . . . , N while N is an integer of 2 or larger;
approximating p(x n ) by using kernel density estimation as
[
Math
.
28
]
p
(
x
)
n
≈
K
(
x
n
)
=
1
N
∑
j
=
1
N
exp
{
-
λ
x
a
-
x
j
2
2
}
where j=1, . . . , N, while λ denotes a bandwidth parameter;
approximating the objective function J θ and a weight w n as
[
Math
.
29
]
J
θ
=
1
∑
n
=
1
N
w
n
∑
n
=
1
N
w
n
A
θ
(
x
n
)
w
n
=
1
K
(
x
n
)
+
ε
where ∈ is a positive constant to avoid zero division; and
using the abnormality estimation model A θ (x) optimized so as to minimize the approximated objective function J θ .
4. The abnormality estimation device according to claim 3 , wherein the weight w n is calculated as
[
Math
.
30
]
w
n
=
exp
(
A
θ
(
x
n
)
max
[
A
θ
(
x
j
)
]
j
)
.
5. A non-transitory computer readable medium that stores a program that causes a computer to function as the abnormality estimation device according to claim 4 .
6. A non-transitory computer readable medium that stores a program that causes a computer to function as the abnormality estimation device according to claim 3 .
7. A non-transitory computer readable medium that stores a program that causes a computer to function as the abnormality estimation device according to claim 2 .
8. A non-transitory computer readable medium that stores a program that causes a computer to function as the abnormality estimation device according to claim 1 .
9. An abnormality estimation device that estimates an anomaly degree of data being input, comprising:
processing circuitry configured to
estimate the anomaly degree of the data, by using an abnormality estimation model that is optimized while using a set of pieces of normal data and is optimized so as to minimize a difference between an anomaly degree of a piece of normal data appearing more frequently and an anomaly degree of a piece of normal data appearing less frequently, wherein the piece of normal data comprises the piece of normal data appearing more frequently and the piece of normal data appearing less frequently, wherein
the abnormality estimation model is trained by applying a larger weight to the piece of normal data appearing less frequently among the set of pieces of normal data than to the piece of normal data appearing more frequently thereby reducing false positives.
10. A non-transitory computer readable medium that stores a program that causes a computer to function as the abnormality estimation device according to claim 9 .
11. An abnormality estimation method for estimating an anomaly degree of an acoustic signal being input, comprising:
a step of estimating the anomaly degree of the acoustic signal, by using an abnormality estimation model that is optimized while using a set of normal sounds and is optimized so as to minimize a difference between an anomaly degree of a normal sound appearing more frequently and an anomaly degree of a normal sound appearing less frequently, wherein the set of normal sounds comprises the normal sound appearing more frequently and the normal sound appearing less frequently, wherein
the abnormality estimation model is trained by applying a larger weight to the normal sound appearing less frequently among the set of normal sounds than to the normal sound appearing more frequently thereby reducing false positives.