Method for creating a non-linear, stationary or dynamic model of a control variable of a machine

• US 20060052991A1
• Filed: 06/29/2005
• Published: 03/09/2006
• Est. Priority Date: 06/30/2004
• Status: Active Grant

• Current Assignee: AVL List GmbH (List Privatstiftung)
• Original Assignee: AVL List GmbH (List Privatstiftung)
• Inventors: Gschweitl, Kurt, Pflugl, Horst, Jakubek, Stefan
• Granted Patent: US 7,444,190 B2
• Time in Patent Office: Days
• US Class Current: 703/8
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: G05B 13/048 using a predictor
• CPC Class Codes: G05B 17/02 electric

Abstract

A method for creating a non-linear, stationary or dynamic overall model of a control variable of a combustion engine or partial systems thereof, is based on simplified partial model functions that are used to determine in a weighted fashion at each desired operating point the total output quantities from the partial model function with an associated weighting function. The difference between the total output quantity and the real value is determined for all real operating points; and in areas of operating points with an absolute value of this difference that is above the preset value, a further model function with a further associated weighting function is used for which the absolute value of the difference stays below the preset value. To use such a method in order to arrive faster, i.e. with fewer iterations, at the optimal overall model that satisfies a statistically substantiated high level of prediction quality and to create an overall model made up of as few partial models as possible, the steps for determining the difference between the total output quantity of the associated partial model functions and a real value of the control value as well as the application of a further model and weighting function are executed as many times as needed until the statistically evaluated prediction quality of the overall model has reached a desired value.

Claims

1. Method for creating a non-linear, stationary or dynamic model of a machine, in particular of a combustion engine or of a partial system thereof, over the entire area of all operating points of the machine, preferably using neuronal nets, and whereby the control variable is dependent on a set of input quantities, for example system parameters, system parameters differentiated by time, time-delayed output quantities with feedback and/or preset values at the respective operating point, and whereby the control variable for a group of operating points is determined by way of a measurement technique from the total space of operating points of the machine, an output quantity is determined per model function at these operating points using a simplified partial model function, and whereby at any operating point the output quantities of each partial model function are added in a weighted fashion to an associated weighting function to arrive at a total output quantity for the respective operating point, and whereby for all operating points with control variable determined by a measurement technique the difference, respectively, between the total output quantity and the value of the control value arrived at by a measurement technique is determined and in areas of operating points where the absolute value of this difference is above a preset value, a further model function with a further associated weighting function is used for which the absolute value of the difference stays below the preset value wherein the step of determining the difference between the total output quantity of the associated partial model functions and a value of the control value arrived at this operating point, that is evaluated by a measurement technique, and the step of the application of a further model function and further weighting function is repeated as many times as needed until the statistically evaluated prediction quality of the overall model has reached the desired value.

2. Method as claimed in claim 1 wherein the respectively most recent step is reversed and the run-through is terminated when the statistically evaluated prediction quality starts to deteriorate again.

3. Method as claimed in claim 1, wherein each weighting function is a radial-symmetrical function in the n-dimensional space of the operating points, and whereby the iso-curves of the weighting function represent a closed geometrical construct (or configuration) in the space of the operating points, and each iso-curve with weighting values of a defined limit, preferably 90% of the maximum value of the weighting function, geometrically encloses a group of real operating points.

4. Method as claimed in claim 3 wherein the iso-curve of the weighting function is a hyper-ellipsoid in the n-dimensional space whose main axes in the space of the operating points have, in principle, any orientation in relation to the system of coordinates.

5. Method as claimed in claim 3, wherein the weighting function is an exponential function of at least one input quantity at the respective operating point.

6. Method as claimed in claim 3, wherein a first weighting function is envisioned whose iso-curve of a defined weighted value geometrically encloses all real operating points.

7. Method as claimed in claim 6 wherein, on the basis of the operating point that has the largest absolute value of the difference between total output quantity of the model and the actual value of the control variable, further points are established for the next partial model in whose center (center of mass) the further weighting function is placed.

8. Method as claimed in claim 3, wherein the number of real operating points within the iso-curve of a defined weighted value of the weighting function is larger than the number or regressors, preferably by at least a factor 1.05.

9. Method as claimed in claim 8 wherein, if the number of regressors within the closed iso-curve of a defined weighted value of the weighting function drops below the number of input quantities that is multiplied by a factor, preferably below the number multiplied by a factor 1.05, this model and the associated weighting function are removed from the overall model.

10. Method as claimed in claim 1, wherein the error of the model function within the defined iso-curves of each weighting function is smaller than 10%.