Computer-readable recording medium, abnormality determination method, and abnormality determination device
A learning device performs learning by an autoencoder, using waveform data with changes over time that is obtained from an intrinsic movement of an object. The learning device performs persistent homology conversion to calculate a change in the number of connected component in accordance with a change of a threshold for a value of the first waveform data. The learning device determines abnormality based on a determination result of a learner to which output data of the autoencoder obtained from the waveform data and output data obtained from the persistent homology conversion are input, and in which machine learning about abnormality of the waveform data has been performed.
1. A non-transitory computer-readable recording medium having stored therein a program that causes a computer to execute a process comprising:
acquiring first data from a trained autoencoder by inputting, into the autoencoder, first waveform data of waveforms that have variations among individuals;
generating second data by applying persistent homology conversion to the first waveform data in accordance with a change of a threshold value for a value of the first waveform data;
training a machine learning model by inputting the first waveform data, the first data and the second data into the machine learning model;
wherein the applying persistent homology conversion to the first waveform data includes:
a process of first counting the number of connected components, which are zero-dimensional holes by decreasing the threshold value uniformly for the first waveform data, and
outputting series data, indicating a transition of a number of connected components, based on a result of the process of first counting; and
a process of second counting the number of connected components, which are zero-dimensional holes by increasing the threshold value uniformly for the first waveform data, and
outputting series data, indicating a transition of a number of connected components, based on a result of the process of second counting, and
the machine learning model is used for a process of determining abnormality for second waveform data based on an output result acquired by inputting second waveform data, third data and fourth data into the machine learning model, the third data being acquired by inputting the second waveform data into the autoencoder, the fourth data being generated by applying persistent homology conversion to the second waveform data.
2. The non-transitory computer-readable recording medium according to claim 1 , wherein
the process of determining includes further input of any one of the third data and the fourth data and a calculation result obtained by performing fast Fourier transform on the first waveform data to a learner, and determine abnormality based on a determination result output by the learner.
3. The non-transitory computer-readable recording medium according to claim 1 , wherein
the process of determining includes converting a result of the persistent homology conversion into a Betti series to input to a learner, and determining abnormality based on a determination result output by the learner.
4. The non-transitory computer-readable recording medium according to claim 1 , wherein the process of determining is implemented by a learner which is a convolutional neural network.
5. The non-transitory computer-readable recording medium according to claim 1 , wherein
the acquiring the first data from the autoencoder occurs by unsupervised learning with normal waveform data.
6. The non-transitory computer-readable recording medium according to claim 1 , the process further comprising:
performing machine learning of a learner by supervised learning with the first waveform data that is learning data to which a label indicating either one of normal and abnormal is added.
7. An abnormality determination method comprising:
acquiring first data from a trained autoencoder by inputting, into the autoencoder, first waveform data of waveforms that have variations among individuals;
generating second data by applying persistent homology conversion to the first waveform data in accordance with a change of a threshold value for a value of the first waveform data;
training a machine learning model by inputting the first waveform data, the first data and the second data into the machine learning model;
wherein the applying persistent homology conversion to the first waveform data includes:
a process of first counting the number of connected components, which are zero-dimensional holes by decreasing the threshold value uniformly for the first waveform data, and
outputting series data, indicating a transition of a number of connected components, based on a result of the process of first counting; and
a process of second counting the number of connected components, which are zero-dimensional holes by increasing the threshold value uniformly for the first waveform data, and
outputting series data, indicating a transition of a number of connected components, based on a result of the process of second counting, and
the machine learning model is used for a process of determining abnormality for second waveform data based on an output result acquired by inputting second waveform data, third data and fourth data into the machine learning model, the third data being acquired by inputting the second waveform data into the autoencoder, the fourth data being generated by applying persistent homology conversion to the second waveform data.
8. The abnormality determination method according to claim 7 , wherein
the process of determining includes further input of any one of the third data and the fourth data and a calculation result obtained by performing fast Fourier transform on the first waveform data to a learner, and determine abnormality based on a determination result output by the learner.
9. The abnormality determination method according to claim 7 , wherein
the process of determining includes converting a result of the persistent homology conversion into a Betti series to input to a learner, and determining abnormality based on a determination result output by the learner.
10. The abnormality determination method according to claim 7 , wherein the process of determining is implemented by a learner which is a convolutional neural network.
11. The abnormality determination method according to claim 7 , wherein
the acquiring the first data from the autoencoder occurs by unsupervised learning with normal waveform data.
12. The abnormality determination method according to claim 7 , further comprising:
performing machine learning of a learner by supervised learning with the first waveform data that is learning data to which a label indicating either one of normal and abnormal is added.
13. The non-transitory computer-readable recording medium according to claim 1 , wherein
the first waveform data is an electrocardiographic waveform that indicates normality and abnormality of heartbeat of an object among the plurality of objects.