CALCULATION DEVICE AND CALCULATION METHOD FOR NEURAL NETWORK

WIPO Patent Application WO/2023/233858

A1

Provided is a calculation device that is for a neural network and that is capable of performing modeling on a target to be modeled by using a mathematical formula based on physical law. A calculation device (1) according to the present invention comprises: a structural optimization unit (16) which is connected to an apparatus (3) having installed therein first and second sensors (31) and which receives input of a detection value of the first sensor (31); and a structural mathematical formula conversion unit (14) for deriving, using a neural network (NN), a prediction formula that is a mathematical formula indicating a relationship between the detection value of the first sensor (31) and a prediction value of a detection value of the second sensor (31). The NN includes a node having a value indicated by a mathematical formula including an exponentiation in which an output value of the node is used as the base and a weighting coefficient set to an edge connected to the node is used as the exponent portion. The structural optimization unit (16) calculates the weighting coefficient and a structure of the NN, from the detection value of the first sensor (31). The structural mathematical formula conversion unit (14) derives a mathematical formula having a term including an exponentiation as a prediction formula, using the weighting coefficient and the structure of the NN calculated by the structural optimization unit (16).

HIRANO MAKOTO (JP)
TOKUDA YUYA (JP)

PCT/JP2023/015901

December 07, 2023

April 21, 2023

Click for automatic bibliography generation  

HITACHI LTD (JP)

G06N3/09; G05B23/02

JPH10134018A	1998-05-22
CN114036582A	2022-02-11

POLAIRE I.P.C. (JP)

View/Download PDF      PDF Help

Previous Patent: CONTROL DEVICE, CONTROL METHOD, AND PROGRAM

Next Patent: TOOL CHANGER, COUPLING UNIT, AND MODULE UNIT