Data generation device and method, and learning device and method

A learning device for training a neural network in which a plurality of layers are connected by a plurality of connection weights, in a flow reaction process for producing a product by reacting a raw material while the raw material is flowing, for quick reaction process, the neural network outputting a production result corresponding to a process condition in a case where the process condition including flow velocity of a raw material is input, the learning device comprising a processor configured to:

• perform a first computation for deriving the connection weight of the neural network by using a data set consisting of a plurality of pieces of learning data consisting of each of the plurality of process conditions and the production result corresponding to the process condition;
• perform a second computation for calculating a prediction result for each process condition of a condition data set having a plurality of process conditions whose production results are unknown using the neural network in which the connection weight is computed to generate a prediction data set in which the process condition and the prediction result are associated with each other, specifying the prediction result closest to a preset target result among a plurality of the obtained prediction results, and extracting a process condition associated with the specified prediction result as an extracted process condition;
• and
• assuming that the number of the data sets is Nj and the number of the connection weights of the neural network is Mj, determine whether or not a degree of difference between the production result in a case where the product is produced under the extracted process condition and the prediction result associated with the extracted process condition is within a preset allowable range, and in a case where the determination is negative, update the number Nj of the data sets by adding, to the data set, production information in which the extracted process condition is associated with the production result in a case where the production is performed under the extracted process condition as the learning data, and update the number Mj of new connection weights of the neural network so that Mj−1≤Mj<Nj/2, and in a case where the determination is positive, determine that the extracted process condition is a process condition used for the process,
• wherein the first computation, the second computation, and the determination are repeated until the determination is positive based on the updated data set and the neural network whose connection weights are updated.

A learning method for training a neural network in which a plurality of layers are connected by a plurality of connection weights, the neural network outputting a production result corresponding to a process condition in a case where the process condition is input in a process for producing a product, the learning method comprising:

• performing a first computation for deriving the connection weight of the neural network by using a data set generated by a data generation method, in a flow reaction process for producing the product by reacting a raw material while the raw material is flowing, for quick reaction process, for generating the data set consisting of a plurality of pieces of learning data for training a neural network in which a plurality of layers are connected by a plurality of connection weights, the neural network outputting a production result corresponding to a process condition in a case where the process condition including flow velocity of the raw material is input for producing the product, the data generation method comprising:
• setting an upper limit and a lower limit of the process condition;
• regarding the flow velocity, assuming that a total number of the connection weights of the neural network is M0, setting a plurality of the process conditions of 2×M0 or more in the range of the upper limit and the lower limit;
• acquiring a production result corresponding to each of the plurality of process conditions, which is derived by producing the product under each of the plurality of process conditions;
• and
• generating the plurality of pieces of learning data consisting of each of the plurality of process conditions and the production result corresponding to the process condition as the data set;
• performing a second computation for calculating a prediction result for each process condition of a condition data set having a plurality of process conditions whose production results are unknown using the neural network in which the connection weight is computed to generate a prediction data set in which the process condition and the prediction result are associated with each other;
• specifying the prediction result closest to a preset target result among a plurality of the obtained prediction results, and extracting a process condition associated with the specified prediction result as an extracted process condition;
• and
• assuming that the number of the data sets is Nj and the number of the connection weights of the neural network is Mj, determining whether or not a degree of difference between the production result in a case where the product is produced under the extracted process condition and the prediction result associated with the extracted process condition is within a preset allowable range, and in a case where the determination is negative, updating the number Nj of the data sets by adding, to the data set, production information in which the extracted process condition is associated with the production result in a case where the production is performed under the extracted process condition as the learning data, and updating the number Mj of new connection weights of the neural network so that Mj−1≤Mj<Nj/2, and in a case where the determination is positive, determining that the extracted process condition is a process condition used for the process,
• wherein the first computation, the second computation, and the determination are repeated until the determination is positive based on the updated data set and the neural network whose connection weights are updated.

A learning method for training a neural network in which a plurality of layers are connected by a plurality of connection weights, in a flow reaction process for producing a product by reacting a raw material while the raw material is flowing, for quick reaction process, the neural network outputting a production result corresponding to a process condition in a case where the process condition including flow velocity of a raw material is input, the learning method comprising:

• performing a first computation for deriving the connection weight of the neural network by using a data set consisting of a plurality of pieces of learning data consisting of each of the plurality of process conditions and the production result corresponding to the process condition;
• performing a second computation for calculating a prediction result for each process condition of a condition data set having a plurality of process conditions whose production results are unknown using the neural network in which the connection weight is computed to generate a prediction data set in which the process condition and the prediction result are associated with each other, specifying the prediction result closest to a preset target result among a plurality of the obtained prediction results, and extracting a process condition associated with the specified prediction result as an extracted process condition;
• and
• assuming that the number of the data sets is Nj and the number of the connection weights of the neural network is Mj, determining whether or not a degree of difference between the production result in a case where the product is produced under the extracted process condition and the prediction result associated with the extracted process condition is within a preset allowable range, and in a case where the determination is negative, updating the number Nj of the data sets by adding, to the data set, production information in which the extracted process condition is associated with the production result in a case where the production is performed under the extracted process condition as the learning data, and updating the number Mj of new connection weights of the neural network so that Mj−1≤Mj<Nj/2, and in a case where the determination is positive, determining that the extracted process condition is a process condition used for the process,
• wherein the first computation, the second computation, and the determination are repeated until the determination is positive based on the updated data set and the neural network whose connection weights are updated.