Physical model identification system
A physical model identification system according to the present invention first uses actual data sampled from a manufacturing line to learn a statistical model expressing a controlled object in the manufacturing line. Next, this system creates numerical quantity data quantifying an input-output relationship between a registered input variable and a response variable of a learned statistical model. Next, this system identifies a correction coefficient of a physical model such that a relationship, expressed by the numerical quantity data, between the registered input variable and the response variable is maintained. An identified correction coefficient is reflected in the physical model implemented in a calculator controlling the manufacturing line. Accordingly, prediction precision of preset values and control values of the controlled object by a practical model is improved, and a stable operation of the manufacturing line and high quality production become possible.
1. A physical model identification system comprising:
a calculator calculating a preset value and a control value, the preset value and the control value being applied to a controlled object, by using a physical model expressing a physical phenomenon in the controlled object including a manufacturing facility and a manufacturing process by a mathematical formula;
a data sampling device sampling plural sets of process data as control results of the controlled object based on the preset value and the control value, the preset value and the control value being calculated by the calculator;
a physical model identification device identifying the physical model based on the plural sets of process data; and
a controller that applies the physical model identified by the physical model identification device to control operations of the manufacturing facility or the manufacturing process performed by the manufacturing facility;
wherein the physical model expresses an input-output relationship between plural input variables and at least one response variable by a mathematical formula including a correction coefficient,
wherein each of the plural sets of process data includes values of the plural input variables and values of the response variable, and
wherein the physical model identification device has:
a data edition unit separating remaining data resulting from removal of abnormal data from the plural sets of sampled process data into data for learning and data for verification;
a statistical model learning unit using the data for learning to inductively learn a statistical model expressing the input-output relationship between the plural input variables and the response variable for the controlled object;
a model validity verification unit verifying validity of the learned statistical model by using the data for verification;
a relationship specifying unit using the learned statistical model to create numerical quantity data quantifying an input-output relationship between at least one registered input variable included in the plural input variables and the response variable in a case where the learned statistical model is valid;
a model coefficient identification unit identifying the correction coefficient of the physical model such that calculated values match values of the response variable which are included in the numerical quantity data, the calculated values being outputs of the physical model in a case where values of the registered input variable which are included in the numerical quantity data are set as inputs to the physical model; and
a model coefficient setting unit setting the identified correction coefficient for the physical model implemented in the calculator.
2. The physical model identification system according to claim 1 , wherein the statistical model is a model expressing a function showing the input-output relationship between the plural input variables and the response variable by using a method of least squares or a model expressing the input-output relationship by using machine learning.
3. The physical model identification system according to claim 1 , wherein
the abnormal data are process data in which a value of at least one of the input variables does not fluctuate among the plural sets of continuous process data or are process data in which at least one input variable fluctuates a threshold value or more among the plural sets of continuous process data.
4. The physical model identification system according to claim 1 , wherein
the learned statistical model is determined to be valid in a case where a difference between
an error A between a calculated value of the response variable, the calculated value being output in a case where a value of the input variable, the value being included in the data for learning, is input to the learned statistical model, and an actual value of the response variable, the actual value being included in the data for learning and
an error B between a calculated value of the response variable, the calculated value being output in a case where a value of the input variable, the value being included in the data for verification, is input to the learned statistical model, and an actual value of the response variable, the actual value being included in the data for verification
is within a threshold value.
5. The physical model identification system according to claim 1 , wherein
the numerical quantity data are a numerical value table describing a relationship between the registered input variable and the response variable or a function approximating the relationship.