Nonlinear neural network fault detection system and method

• US 20070288409A1
• Filed: 06/24/2005
• Published: 12/13/2007
• Est. Priority Date: 05/31/2005
• Status: Active Grant

• Current Assignee: Honeywell International Inc.
• Original Assignee: Honeywell International Inc.
• Inventors: Mylaraswamy, Dinkar, Menon, Sunil, Mukherjee, Joydeb, Kini, Venkataramana
• Granted Patent: US 7,660,774 B2
• Time in Patent Office: Days
• US Class Current: 706/25
• CPC Class Codes: G06N 3/045 Combinations of networks

Abstract

A system and method for fault detection is provided. The fault detection system provides the ability to detect symptoms of fault in turbine engines and other mechanical systems that have nonlinear relationships. The fault detection system uses a neural network to perform a data representation and feature extraction where the extracted features are analogous to principal components derived in a principal component analysis. This neural network data representation analysis can then be used to determine the likelihood of a fault in the system.

Claims

1. A fault detection system for detecting faults in a turbine engine, the fault detection system comprising:

an encoding neural network, the encoding neural network receiving sensor data from the turbine engine, the encoding neural network creating plurality of scores from the sensor data, the plurality of scores comprising a reduced feature space representation of the sensor data; and

a decoding neural network, the decoding neural network receiving the plurality of scores and creating a reconstructed estimate of the sensor data.

2. The system of claim 1 further comprising a classifier, the classifier comparing the sensor data to the reconstructed estimate of the sensor data to determine a likelihood of a fault in the turbine engine.

3. The system of claim 1 further comprising a classifier, the classifier comparing the plurality of scores to a set of scores generated from historical sensor data.

4. The system of claim 1 wherein the encoding neural network comprises a plurality of nodes, and wherein at least one of the plurality of nodes includes a nonlinear activation function for modeling nonlinear relationships.

5. The system of claim 1 wherein the encoding neural network comprises a mapping layer, the mapping layer including a plurality of nodes, and wherein the plurality of nodes in the mapping layer each includes a nonlinear activation function for modeling nonlinear relationships.

6. The system of claim 5 wherein the encoding neural network further comprises an output layer, the output layer including a plurality of nodes, and wherein the plurality of nodes in the output layer each includes a linear activation function.

7. The system of claim 1 wherein the encoding neural network comprises a multilayer neural network trained using a gradient descent method using an objective function.

8. The system of claim 1 wherein the encoding neural network is trained with an objective function and historical sensor data.

9. The system of claim 8 wherein the objective function is defined as:

$E_{\mathrm{SANLPCA}}=\max <br><br>\underset{i=1}{\overset{p}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>t_i^T<br><br>t_i+\min <br><br>\underset{1\ne <br><br>j}{\overset{p}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br><br><br>t_i^T<br><br>t_j<br><br>$

where t_i is a vector of an i^th score generated from an output of the encoding neural network using the historical sensor data as an input to the encoding neural network.

10. The system of claim 8 wherein the objective function is selected to maximize variance components and minimize covariance components.

11. The system of claim 8 wherein the objective function includes a sum of variance components and a sum of covariance components, and wherein the encoding neural network is trained by maximizing the sum of variance components and minimizing the sum of covariance components.

12. The system of claim 1 wherein the decoding neural network is trained using a supervised mode of training using historical sensor data, scores obtained from the output of the encoding neural network as inputs to the decoding neural network and the historical sensor data as target output values.

13. A method of training an encoding neural network for fault detection in a system, the method comprising the steps of:

providing an objective function, the objective function including variance components and covariance components of scores;

providing a set of historical sensor data as input to the encoding neural network; and

optimizing weight assignments in the encoding neural network to maximize the variance components and minimize the covariance components of scores obtained using the set of historical sensor data.

14. The method of claim 13 wherein the objective function includes a weighted sum of the variance components and a weighted sum of the covariance components.

15. The method of claim 13 wherein the objective function is defined as:

$E_{\mathrm{SANLPCA}}=\max <br><br>\underset{i=1}{\overset{p}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>t_i^T<br><br>t_i+\min <br><br>\underset{1\ne <br><br>j}{\overset{p}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br><br><br>t_i^T<br><br>t_j<br><br>$

where t_i is a vector of an i^th score generated from an output of the encoding neural network using the historical sensor data as the input to the encoding neural network.

16. The method of claim 13 wherein the encoding neural network comprises a mapping layer, the mapping layer including a plurality of nodes, and wherein the plurality of nodes in the mapping layer each includes a nonlinear activation function for modeling nonlinear relationships.

17. The method of claim 16 wherein the encoding neural network further comprises an output layer, the output layer including a plurality of nodes, and wherein the plurality of nodes in the output layer each include a linear activation function.

18. The method of claim 13 wherein the encoding neural network creates a plurality of scores from sensor data, the plurality of scores comprising a reduced feature space representation of the sensor data.

19. The method of claim 13 further comprising the step of training a decoding neural network to receive a plurality of scores from the encoding neural network output and create a reconstructed estimate of sensor data.

20. The method of claim 19 where the step of training the decoding neural network comprises using a supervised mode of training using historical sensor data, scores obtained from the output of the encoding neural network as inputs to the decoding neural network and the historical sensor data as target output values.

21. A program product comprising:

a) a fault detection program, the fault detection program including;

an encoding neural network, the encoding neural network receiving sensor data from a turbine engine, the encoding neural network creating plurality of scores from the sensor data, the plurality of scores comprising a reduced feature space representation of the sensor data; and

a decoding neural network, the decoding neural network receiving the plurality of scores and creating a reconstructed estimate of the sensor data; and

b) computer-readable signal bearing media bearing said program.