Abnormality detection device, learning device, abnormality detection method, and learning method

1. An abnormality detection device comprising:

a computer configured to operate as:

an encoder configured to compress input data using a compressing parameter to generate compressed data;

a first identifier configured to detect a first abnormality of the input data on the basis of a logarithm probability density on a prior distribution of the compressed data and a logarithm density ratio of the compressed data distribution to the prior distribution;

a decoder configured to decode compressed data using a decoding parameter to generate reconstructed data; and

a second identifier configured to calculate a difference between the reconstruction data and the input data, and to detect a second abnormality of the input data on the basis of the difference,

wherein the first identifier compares intermediate layer data of compressed data compressed by the encoder with intermediate layer data of the prior distribution and detects the first abnormality, and

the second identifier compares intermediate layer data of reconstruction data decoded by the decoder with intermediate layer data of the input data and detects the second abnormality.

2. The abnormality detection device according to claim 1 ,

wherein the first identifier detects that the input data is the first abnormality when a sum of a logarithm probability density on the prior distribution of compressed data compressed by the encoder and a logarithm density ratio of a compressed data distribution to a prior distribution is equal to or less than a first threshold value, and

the second identifier detects that the input data is the second abnormality when a difference between reconstruction data decoded by the decoder and the input data is equal to or greater than a second threshold value.

3. A learning device comprising:

a computer configured to operate as:

an encoder configured to compress input data using a compression parameter to generate a compressed data;

a first identifier configured to determine whether a distribution of the compressed data input by the encoder is a distribution of the compressed data or a prior distribution prepared in advance, and input a first identification result to the encoder;

a decoder configured to decode the compressed data using a compressing parameter to generate reconstructed data; and

a second identifier configured to determine whether the reconstructed data input by the decoder is the reconstruction data or the input data and output a second identification result to the encoder and the decoder,

wherein the encoder adjusts the compressing parameter on the basis of the first identification result and the second identification result, and

the decoder adjusts the decoding parameter on the basis of the second identification result,

wherein the encoder adjusts the compressing parameter to reduce a difference between the distribution of compressed data and the prior distribution, and

the decoder adjusts the decoding parameter to reduce a difference between the reconstruction data and learning data.

4. A learning device comprising:

a computer configured to operate as:

an encoder configured to compress input data using a compression parameter to generate a compressed data;

a first identifier configured to determine whether a distribution of the compressed data input by the encoder is a distribution of the compressed data or a prior distribution prepared in advance, and input a first identification result to the encoder;

a decoder configured to decode the compressed data using a compressing parameter to generate reconstructed data; and

a second identifier configured to determine whether the reconstructed data input by the decoder is the reconstruction data or the input data and output a second identification result to the encoder and the decoder,

wherein the encoder adjusts the compressing parameter on tale basis of the first identification result and the second identification result, and

the decoder adjusts the decoding parameter on the basis of the second identification result

wherein the first identifier compares intermediate layer data of compressed data compressed by the encoder with intermediate layer data of the prior distribution and performs the identification, and

the second identifier compares intermediate layer data of reconstruction data decoded by the decoder with intermediate layer data of learning data and performs the identification.

5. An abnormality detection method comprising:

compressing input data using a compressing parameter to generate a compressed data;

detecting a first abnormality of the input data on the basis of a logarithm probability density on a prior distribution of the compressed data and a logarithm density ratio of the compressed data distribution to the prior distribution;

decoding the compressed data using a decoding parameter to generate a reconstructed data; and

calculating a difference between the reconstruction data and the input data, and detecting a second abnormality of the input data on the basis of the difference,

wherein the detecting the first abnormality compares intermediate layer data of compressed data compressed by the compressing with intermediate layer data of the prior distribution and detects the first abnormality, and

the calculating the difference compares intermediate layer data of reconstruction data decoded by the decoding with intermediate layer data of the input data and detects the second abnormality.

6. A learning method comprising:

compressing input data using a compression parameter to generate compressed data;

identifying whether a distribution of the compressed data input by an encoder is a distribution of the compressed data or a prior distribution prepared in advance, and inputting a first identification result to the encoder;

decoding the compressed data using a compressing parameter to generate a reconstructed data;

identifying whether the reconstructed data input by a decoder is the reconstruction data or the input data and output a second identification result to the encoder and the decoder;

adjusting the compressing parameter on the basis of the first identification result and the second identification result; and

adjusting the decoding parameter on the basis of the second identification result,

wherein the compressing adjusts the compressing parameter to reduce a difference between the distribution of compressed data and the prior distribution, and

the decoding adjusts the decoding parameter to reduce a difference between the reconstruction data and learning data.

7. A learning method comprising:

compressing input data using a compression parameter to generate compressed data;

first identifying whether a distribution of the compressed data input by an encoder is a distribution of the compressed data or a prior distribution prepared in advance, and inputting a first identification result to the encoder;

decoding the compressed data using a compressing parameter to generate a reconstructed data;

second identifying whether the reconstructed data input by a decoder is the reconstruction data or the input data and output a second identification result to the encoder and the decoder;

first adjusting the compressing parameter on the basis of the first identification result and the second identification result; and

second adjusting the decoding parameter on the basis of the second identification result,

wherein the first identifying compares intermediate layer data of compressed data compressed by the compressing with intermediate layer data of the prior distribution and performs the identification, and

the second identifying compares intermediate layer data of reconstruction data decoded by the decoding with intermediate layer data of learning data and performs the identification.