Method and control device for controlling a process including the use of a neural network having variable network parameters

A method for controlling a process in a controlled system having at least one process sequence, comprising:

• supplying a plurality of input variables to a neural network having variable network parameters for precalculating a process parameter before the start of a process sequence, said process parameter being a rolling force on a rolling stand in a mill train and said input variables include a relative decrease in thickness of a rolling stock in the stand, an entrance temperature of the rolling stock, a tension in the rolling stock in front of the stand, a tension in the rolling stock behind the stand, a radius of the rolling stock, a width of a roll in the roll stand, and a thickness of the rolling stock in front of the stand, said input variables define a multi-dimensional input space according to their number, said input variables are linked by linear linkage of Gaussian functions to form the process parameter to be precalculated, each of said Gaussian functions are multi-dimensional corresponding to a dimension of the input space, each of said Gaussian functions has a variable center, width, and amplitude;
• presetting a controlled system at the start of each process sequence as a function of said precalculated process parameter;
• measuring values for said input variables and said process parameter during said process sequence;
• supplying said measured input variable values after said process sequence to said neural network;
• calculating in said neural network a network response based on the measured input variables;
• calculating a deviation between said network response and said measured process parameter;
• adapting said network parameters in said neural network to reduce said deviation, said adapting step takes place on-line, such that after each measuring step the measured input variable values and process parameter obtained are used for adaptation of the network parameters;
• determining corresponding values for said process parameters based on physical model assumptions, prior to the first process sequence, from randomly preset values for the input variables;
• preadapting said neural network with said corresponding values
• shifting one Gaussian function upon each adaptation step whose center is closest to the point defined in the input space by the measured input variables, said shifting step taking place by a preset step size in a direction toward said point;
• and
• adapting the width of the Gaussian functions and the amplitude of the shifted Gaussian function with preset step sizes as a function of, and in order to reduce, a deviation between the network response generated by the network with the shifted Gaussian function for the measured input variables and the measured process parameter.

A control device for controlling a process in a controlled system having at least one process sequence, comprising:

• a device which presets the system as a function of a precalculated process parameter which exhibits a system-induced dependence on input variables, said input variables define a multi-dimensional input space according to their number, said input variables are linked by linear linkage of Gaussian functions to form the process parameter to be precalculated, each of said Gaussian functions are multi-dimensional corresponding to a dimension of the input space, each of said Gaussian functions has a variable center, width, and amplitude, such that one Gaussian function upon each adaptation step is shifted whose center is closest to the point defined in the input space by the measured input variables, said shifting step taking place by a preset step size in a direction toward said point, such that the width of the Gaussian functions and the amplitude of the shifted Gaussian function are adapted with preset step sizes as a function of, and in order to reduce, a deviation between the network response generated by the network with the shifted Gaussian function for the measured input variables and the measured process parameter;
• a neural network having variable network parameters for adaptive simulation of an interdependence between the input variables and the process parameter before the start of a process sequence, values for said input variables and said process parameter are measured during said process sequence, such that said measured input variable values are supplied to said neural network after said process sequence, wherein a deviation between a network response based on the measured input variables and the measured process parameter is calculated and said network parameters in said neural network are adapted to reduce said deviation, said network parameters are adapted on-line, such that after each measuring step the measured input variable values and process parameter obtained are used for adaptation of the network parameters, wherein corresponding values for said process parameters based on physical model assumptions are determined, prior to the first process sequence, from randomly preset values for the input variables, and said neural network is preadapted with said corresponding values;
• and
• a mill train including at least one rolling stand, such that said process parameter is a rolling force on said at least one rolling stand in said mill train, wherein for precalculating the rolling force on said at least one roll stand, said input variables include a relative decrease in thickness of a rolling stock in said at least one stand, an entrance temperature of the rolling stock, a tension in the rolling stock in front of said stand, a tension in the rolling stock behind said stand, a radius of the rolling stock, a width of a roll in said roll stand, and a thickness of the rolling stock in front of said stand.