Method and computer program for analyzing variables using pruning, influencing a combustion process in a combustion chamber, using a trainable statistical model

• US 7,945,523 B2
• Filed: 09/27/2004
• Issued: 05/17/2011
• Est. Priority Date: 09/30/2003
• Status: Active Grant

• Current Assignee: Siemens Energy Global GmbH & Co. KG (Siemens AG)
• Original Assignee: Siemens AG
• Inventors: Sterzing, Volkmar, Hofmann, Reimar
• Primary Examiner(s): Vincent; David R
• Application Number:
  US10/574,171
• Publication Number:
  US 20090240636A1
• Time in Patent Office: 2,423 Days
• Field of Search: 706/12, 706/15, 706/16, 706/21, 436/106, 436/116, 436/118, 701 99-102
• US Class Current: 706/12
• CPC Class Codes: G05B 17/02 electric

Abstract

The invention relates to sensitivity analysis of variables influencing a combustion process. A trainable, statistical model is trained in such a way that it describes the combustion process in the combustion chamber. The trained statistical model is used to determine the influence of the variables on said combustion process in the combustion chamber.

Claims

1. A method for analyzing variables influencing a combustion process in a combustion chamber, comprising:

training a statistical model using predetermined values of influencing variables;

altering parameters of the model such that the model describes the combustion process; and

determining the influence of the altered parameters on the combustion process,

wherein the model parameters are weights of the model,

wherein the variables are analyzed by a pruning method and a pruning of the model parameters is performed during the pruning method, and

wherein the pruning method comprising;

representing each of the influencing variables by a node,

providing a link between each of the nodes,

determining a zero order statistical independence and dependence between two of the nodes,

removing the link between the two nodes when a zero order statistical independence is determined between the two of the nodes;

removing a link between a first node and a second node based on a first order statistical independence such that the link between the first node and the second node is removed when the first node is linked to a third node and the second node is linked to the third node.

2. The method according to claim 1, wherein the influencing variables are selected from the group consisting of:

external air pressure, gas pressure, pressure after compression, pressure difference at air filter, pilot gas, load, air temperature, gas temperature, temperature after compression, pressure reduction in a combustion chamber, blade position, gas turbine output, air temperature inflow, air pressure, pressure difference over air filter, pressure after compression stage, temperature after compression stage, inlet guide blade position, pressure difference in combustion chamber, pilot gas stream, gas pressure, gas temperature, rotational speed, and exhaust gas temperature.

3. The method according to claim 2, wherein the combustion process is represented by a variable that describes combustion chamber humming, and the variables that influence the combustion chamber humming variable are graphically represented.

4. The method according to claim 3, wherein the combustion process is represented by a variable that describes NOx emission, and the variables that influence the NOx emissions are graphically represented.

5. The method according to claim 4, wherein the combustion chamber humming variable is a target variable during training.

6. The method according to claim 5, wherein the variable describing the NOx emission is the target variable during training.

7. The method according to claim 6, wherein the trainable statistical model is a neural network, a causal network or a decision tree.

8. The method according to claim 7, wherein the neural network is a feed forward neural network.

9. The method according to claim 1 wherein the combustion chamber is a combustion chamber of a gas turbine.

10. The method according to claim 9, wherein a sensitivity analysis for influencing variables of a combustion process of a combustion chamber is performed.

11. The method according to 10, wherein a sequence of significance of variables influencing combustion chamber humming is determined.

12. The method according to claim 11, wherein a sequence of significance of variables influencing NOx emission of a combustion chamber is determined.

13. The method according to claim 1, wherein the pruning method is based on a statistical independence determined from the weights of the model.

14. A computer program for analyzing variables influencing a combustion process in a combustion chamber, the computer program compising a non-transitory computer useable medium including a computer readable program, wherein the computer readable program when executed on a computer causes the computer to perform a method comprising:

computer code that trains a statistical model using predetermined values of influencing variables;

computer code that alters parameters of the model such that the model describes the combustion process; and

computer code that determines the influence of the altered parameters on the combustion process,

wherein the model parameters are weights of the model, and

wherein the variables are analyzed by a pruning method and a pruning of the model parameters is performed during the pruning method,

wherein the pruning method comprising;

representing each of the influencing variables by a node,

providing a link between each of the nodes,

determining a zero order statistical independence and dependence between two of the nodes,

removing the link between the two nodes when a zero order statistical independence is determined between the two of the nodes;

removing a link between a first node and a second node based on a first order statistical independence such that the link between the first node and the second node is removed when the first node is linked to a third node and the second node is linked to the third node.

15. The computer program according to claim 14, wherein the pruning method is based on a statistical independence determined from the weights of the model.