Process for optimizing control parameters for a system having an actual behavior depending on the control parameters

• US 5,598,076 A
• Filed: 06/09/1994
• Issued: 01/28/1997
• Est. Priority Date: 12/09/1991
• Status: Expired due to Fees

• Current Assignee: Siemens AG
• Original Assignee: Siemens AG
• Inventors: Bocionek, Siegfried, Moller, Marcus, Joppich, Martin, Neubauer, Werner
• Primary Examiner(s): Masih, Karen
• Application Number:
  US08/244,756
• Time in Patent Office: 964 Days
• Field of Search: 318/568.22, 318/568.23, 318/568.24, 318/568.13, 318/568.15, 318/565, 318/561, 318/563, 318/569, 318/567, 364/151, 364/157, 364/148, 364/150, 364/474.15, 364/474.12, 364/474.13, 395/900, 395/21, 395/20
• US Class Current: 318/568.22
• CPC Class Codes: B25J 9/161 Hardware, e.g. neural networks, fuzzy logic, interfaces, processor
• CPC Class Codes: G05B 13/027 using neural networks only

Abstract

A process for optimizing the control parameters of the operation of a system, such as a robotic system or heating system, provides that a trainable component or model is taught the actual behavior of the system, the difference or deviation of the actual behavior from a reference behavior of the system is determined and correction values are provided. New control parameters are determined which take into account the correction values. The determination of the deviation and the incorporation of the correction values is continued until the difference of the actual behavior from the reference behavior is below a predetermined value.

Claims

1. A process for generating control parameters for a reference behavior of a system having an actual behavior depending on control parameters, comprising the steps of:

a) teaching a trainable component said actual behavior of said system,

b) determining a difference between said reference behavior and said actual behavior of said system, which is a reference deviation, as a function of said control parameters;

c) supplying said difference to at least one trainable component to which said actual behavior of said system has been taught,

d) providing from said at least one trainable component a correction value for said control parameters corresponding to said reference deviation,

e) determining new control parameters with aid of said correction value in such a way that said control parameters bring about an improvement of said actual behavior with respect to said reference behavior of said system, and

f) performing at least the steps b) through e) until a difference between said reference behavior and said actual behavior of said system falls below a predetermined value.

2. The process as claimed in claim 1, wherein said trainable component is a neural network which is taught, and further comprising the step of:

weighting said at least one control parameter at least one neuron of said neural network for simulating said actual behavior.

3. The process as claimed in claim 2, further comprising the step of:

teaching said neural network linear behavior of said system.

4. The process as claimed in claim 2, further comprising the steps of:

teaching said neural network nonlinear behavior of said system, and

encoding said control parameters supplied for training said system.

5. The process as claimed in claim 2, further comprising the step of:

assigning to at least one neuron a subset of said control parameters known from said actual behavior for learning said reference behavior.

6. The process as claimed in claim 1, wherein said system has at least minimum-phase behavior, wherein said step of teaching said trainable component determines associated control parameters from said actual behavior of said system, and further comprising the step of:

determining a control parameter deviation from said reference deviation which is combined with said control parameters that brought about said reference deviation.

7. The process as claimed in claim 1, further comprising the steps of:

using two identically taught trainable components as models,

providing a first one of said models with substantially identical control parameters as said control parameters supplied to said system,

providing a second one of said models with modified control parameters that are modified as compared to said control parameters supplied to said system, and

obtaining a deviation by comparing actual behaviors produced by said first and second models, said deviation corresponds to a modification of said modified control parameters on said second model compared to said control parameters of said system,

adapting the modification iteratively so that a square of a sum of said reference deviation brought about by said system and a linear combination of said deviation brought about by said first and second models becomes minimal, and

combining the modification of said control parameters found in said adapting step with particular control parameters that brought about said reference deviation of said system.

8. The process as claimed in claim 1, wherein said system comprises a robot arm that guides a tool, and said control parameters are coordinate specifications which are to be included in a trajectory along which said robot arm is to guide said tool, and wherein said actual behavior of said system is described by said trajectory along which said robot arm actually guides said tool.

9. The process as claimed in claim 1, wherein said system comprises a heating system and said control parameters specify a temperature profile with which heated material is to be heated, and wherein said actual behavior is described by an actual temperature waveform of said heated material.

10. The process as claimed in claim 1, further comprising the steps of:

determining said actual behavior of said system by sensors.