US 12,254,018 B2
Time-series data processing method, time-series data processing device, time-series data processing system, and recording medium
Takashi Konashi, Tokyo (JP)
Assigned to NEC CORPORATION, Tokyo (JP)
Appl. No. 18/013,405
Filed by NEC Corporation, Tokyo (JP)
PCT Filed Jul. 3, 2020, PCT No. PCT/JP2020/026287
§ 371(c)(1), (2) Date Dec. 28, 2022,
PCT Pub. No. WO2022/003983, PCT Pub. Date Jan. 6, 2022.
Prior Publication US 2023/0252038 A1, Aug. 10, 2023
Int. Cl. G06F 7/00 (2006.01); G06F 16/2458 (2019.01); G06F 16/248 (2019.01); G06F 16/28 (2019.01)
CPC G06F 16/2477 (2019.01) [G06F 16/248 (2019.01); G06F 16/284 (2019.01)] 11 Claims
OG exemplary drawing
 
1. A time-series data processing method performed by a computer and comprising:
receiving, from a sensor, first time-series data regarding a monitoring object measured by the sensor in a plurality of first periods, the first time-series data including a plurality of first pieces of partial-time series data, wherein the monitoring object is operating in an abnormal state in the first periods;
learning a model that predicts whether the monitoring object is operating in the abnormal state, by respectively extracting first feature amounts from the first pieces of partial time-series data, wherein the extracted first feature amounts constitute the learned model;
storing the learned model that is constituted by the extracted first feature amounts;
receiving, from the sensor, second time-series data measured by the sensor in a plurality of second periods respectively corresponding to the plurality of first periods in position relationship on a time axis, the second time-series data including a plurality of second pieces of partial-time series data, wherein whether the monitoring object is operating in the abnormal state in the first periods is not known;
respectively extracting second feature amounts from the second pieces of time-series data;
respectively comparing the extracted first feature amounts of the first pieces of partial time-series data in the first periods, which constitute the learned model, with the second feature amounts of the extracted second pieces of partial time-series data, to determine similarities between the extracted first feature amounts and the extracted second feature amounts;
calculating similarities between the first feature amounts and the second feature amounts compared with respect to a target period and similarities between the first feature amounts and the second feature amounts compared with respect to at least one other period;
tallying the similarities with respect to the target period and the similarities with respect to the at least one other period by different methods;
detecting whether the monitoring object is in the abnormal state based on a result of the tallying; and
outputting a result of comparison of the first feature amounts with the second feature amounts.