Process and arrangement for conditioning an input variable of a neural network

A control arrangement for a chemical process, comprising:

• a) at least one sensor mounted for acquiring measured values of the chemical process;
• b) at least one memory connected to said at least one sensor for storing the measured values sensed by said at least one sensor;
• c) a processor connected to said at least one memory for processing the measured values stored in said at least one memory to produce conditioned input values;
• d) a forwardly directed neural network connected to receive the conditioned input values from said processor;
• and
• e) at least one effector connected to receive an output from said forwardly directed neural network for influencing the chemical process and cause a change in the measured values sensed by said at least one sensor,
• f) said at least one memory being connected to store in the memory a time series that is formed from a set of the measured values measured at regular time intervals, the time series being fed to said processor;
• said processor means being programmed to combine a plurality of the measured values in respective intervals in such a way that a length of the intervals within the time series is selected to be greater the further the measured values located therein lie back in the past, and deriving a single value from each interval by convolution with a bell-shaped curve, and feeding said single value from each interval together with a further value from the time series stored in the memory to the neural network as the conditioned input values,
• said neural network being connected to form from the conditioned input values a manipulated variable for the chemical process, and the manipulated variable being forwarded to the effector to control said chemical process.

A method for controlling a physical process utilizing an output quantity of a neural network, comprising the steps of:

• measuring conditions of the physical process, said conditions being influenced by changes in the physical process;
• collecting a set of values over time from the conditions of the physical process measured in said measuring step, said set of values being input data for the neural network, the set of values being including a variable having a relatively long time constant and a variable having a relatively short time constant;
• forming a time series from said set of values by determining the input variable at discrete times;
• delimiting intervals within the time series in such a way that a length of the intervals is selected to be greater the further the values therein lie back in the past;
• convoluting a first partial time series delimited by the first interval with a bell-shaped function to form an average and forming the first maximum value of the convolution result;
• feeding the first maximum value to the neural network which is representative of the first partial time series with another input value which is located within the time series but outside the first interval;
• detecting an output quantity of the neural network;
• and
• controlling the physical process using the output quantity from the neural network, said controlling step including implementing a physical change to said physical process which results in a change in said predetermined conditions of said physical process being measured in said measuring step.