US 12,307,345 B2
State estimation device, state estimation program, estimation model, and state estimation method
Noriaki Hirokawa, Chiyoda-ku (JP); Yoshitaka Inoue, Chiyoda-ku (JP); and Yusuke Fukazawa, Chiyoda-ku (JP)
Assigned to NTT DOCOMO, INC., Chiyoda-ku (JP)
Appl. No. 17/599,808
Filed by NTT DOCOMO, INC., Chiyoda-ku (JP)
PCT Filed Apr. 2, 2020, PCT No. PCT/JP2020/015253
§ 371(c)(1), (2) Date Sep. 29, 2021,
PCT Pub. No. WO2020/213425, PCT Pub. Date Oct. 22, 2020.
Claims priority of application No. 2019-079942 (JP), filed on Apr. 19, 2019.
Prior Publication US 2022/0198341 A1, Jun. 23, 2022
Int. Cl. G06N 3/04 (2023.01); G06F 18/2431 (2023.01); G06N 3/045 (2023.01); G06N 20/20 (2019.01)
CPC G06N 20/20 (2019.01) [G06F 18/2431 (2023.01); G06N 3/045 (2023.01)] 11 Claims
OG exemplary drawing
 
1. A state estimation device for estimating a user's state based on sensor data detected by a sensor, the state estimation device comprising circuitry configured to:
acquire sensor information composed of time-series sensor data acquired by at least one sensor;
provisionally allocate one of state labels for identifying each of user's states to each time and one or more types of sensor data associated with the time;
calculate a feature parameter indicating a feature of a plurality of sensor data to which a same state label is allocated, and perform learning of a first model that classifies the sensor data associated with each time into any one of a plurality of states;
extract, from a state sequence being a time-series sequence of the state labels allocated to each time, information indicating transition of the state labels before a transition reference time being one time point in the time series and a state label allocated to a time subsequent to the transition reference time as an input feature quantity and a supervision label, respectively, and perform learning of a second model whose input is transition of state labels before the transition reference time and whose output is a state label at a time subsequent to the transition reference time by machine learning based on learning data including the extracted input feature quantity and the extracted supervision label;
estimate a state at each time based on each of the first model and the second model for each time in the time series, and update a state label allocated to each time based on each estimated state; and
output the first model and the second model when specified conditions for update of the state labels are satisfied.