Abnormality degree calculation system and abnormality degree calculation method

An abnormality degree calculation system that calculates an abnormality degree of a target device, the abnormality degree calculation system comprising:

• a concept classification assignment unit configured to assign, based on a type number of the target device, a predetermined concept classification identifying the target device among a plurality of devices belonging to a common model and sharing a same-level concept;
• a feature value vector extraction unit configured to generate, from sensor data obtained from a sensor corresponding to the target device and measuring sound or vibration produced during operation of the target device, a feature value vector representing physical characteristics of the sensor data;
• a likelihood calculation unit configured to calculate, by inputting the feature value vector into a machine learning model trained to model a probability distribution of normal sensor data, a likelihood indicating conformity of the feature value vector to normal operation of the target device,
• wherein the machine learning model is trained such that
• feature value vectors derived from sensor data of the target device corresponding to the assigned concept classification are learned as normal data and output with a high likelihood, and
• feature value vectors derived from sensor data of other devices having the same-level concept as the target device are learned as abnormal data and output with a low likelihood, even when the other devices are operating normally,
• wherein, during training, model parameters are updated so as to enforce asymmetric likelihood behavior by increasing penalties when the likelihood for the target device decreases or when the likelihood for the other devices increases;
• and
• an abnormality degree calculation unit configured to calculate an abnormality degree representing a degree of abnormal operation of the target device, based on a likelihood that the likelihood calculation unit calculates.