METHODS FOR PROGNOSING MECHANICAL SYSTEMS

• US 20100023307A1
• Filed: 07/24/2009
• Published: 01/28/2010
• Est. Priority Date: 07/24/2008
• Status: Active Grant

• Current Assignee: University of Cincinnati
• Original Assignee: University of Cincinnati
• Inventors: Liao, Linxia, Lee, Jay
• Granted Patent: US 8,301,406 B2
• Time in Patent Office: Days
• US Class Current: 703/7
• CPC Class Codes: G05B 23/0254 based on a quantitative model, e.g. mathematical relationships between inputs and outputs; functions: observer, Kalman filter, residual calculation, Neural Networks
• CPC Class Codes: G06F 18/2137 based on criteria of topology preservation, e.g. multidimensional scaling or self-organising maps
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/2321 using statistics or function optimisation, e.g. modelling of probability density functions
• CPC Class Codes: G06F 2218/00 Aspects of pattern recognition specially adapted for signal processing

Abstract

A method of prognosing a mechanical system to predict when a failure may occur is disclosed. Measurement data corresponding to the mechanical system is used to extract one or more features by decomposing the measurement data into a feature space. A prediction model is then selected from a plurality of prediction models for the one or more features based at least on part on a degradation status of the mechanical system and a reinforcement learning model. A predicted feature space is generated by applying the selective prediction model to the feature space as well as a confidence value by comparing the predicted feature space with a normal baseline distribution, a faulty baseline distribution, or a combination thereof. A status of mechanical system based at least in part on the confidence value is then provided.

Claims

1. A method of prognosing a mechanical system comprising:

receiving measurement data corresponding to the mechanical system;

extracting one or more features from the received measurement data by decomposing the measurement data into a feature space;

selecting a prediction model from a plurality of prediction models for one or more features based at least on part on a degradation status of the mechanical system and a reinforcement learning model;

generating a predicted feature space by applying the selective prediction model to the feature space;

generating a confidence value by comparing the predicted feature space with a normal baseline distribution, a faulty baseline distribution, or a combination thereof; and

providing a status of mechanical system based at least in part on the confidence value.

2. A method as claimed in claim 1 wherein the measurement data comprises current data, voltage data, vibration data, pressure data, temperature data, acoustic emissions, or combinations thereof.

3. A method as claimed in claim 1 wherein the one or more features comprises one or more time domain features, one or more frequency domain features, or combinations thereof.

4. A method as claimed in claim 3 wherein the method further comprises obtaining the frequency domain features by applying a Fourier transform to stationary signals within the measurement data, and applying a wavelet packet transform to non-stationary signals within the measurement data.

5. A method as claimed in claim 1 wherein the features are extracted by a time domain analysis, a frequency domain analysis or combinations thereof.

6. A method as claimed in claim 1 wherein the method further comprises dimensionally reducing the feature space to generate a reduced set of uncorrelated features from the features within the feature space.

7. A method as claimed in claim 6 wherein dimensionally reducing the feature space further comprises applying a principal component analysis, a Karhunen-Loé

ve transform, or a combination thereof to the feature space.

8. A method as claimed in claim 1 wherein the degradation status of the mechanical system is determined by comparing the feature space with the normal baseline feature space.

9. A method as claimed in claim 1 wherein the degradation status is based on a performance index generated by a self-organizing map trained with measurement data of a normal operating state.

10. A method as claimed in claim 9 wherein the performance index is the difference between an input vector corresponding with the feature space and a weight vector.

11. A method as claimed in claim 1 wherein the plurality of prediction models comprises one or more auto-regressive moving average models, one or more recurrent neural network models, or combinations thereof.

12. A method as claimed in claim 1 wherein:

the reinforcement learning model is defined by a plurality of states, each state corresponding to a particular degradation status;

the reinforcement learning model comprises a Q-value for each prediction model at each state; and

the selected prediction model is the prediction model having largest Q-value at a particular state.

13. A method as claimed in claim 12 wherein the states are based, at least in part, on a performance index generated by a self-organizing map trained with measurement data of a normal operating state.

14. A method as claimed in claim 13 wherein the Q-values are developed by an iterative learning process.

15. A method as claimed in claim 14 wherein the iterative learning process comprises:

choosing a prediction model in a particular state;

generating a predicted output;

comparing the predicted output with a real value of historical data; and

assigning a reward value to the prediction model in the particular state such that a positive reward value is assigned when predicted output has a relatively high prediction accuracy and a negative reward value is assigned when the predicted output has a relatively low prediction accuracy.

16. A method as claimed in claim 15 wherein the Q-value for each predictive model comprises a summation of a plurality of reward values based on a plurality of prediction outputs at the particular state.

17. A method as claimed in claim 1 wherein the predictive feature space is approximated by a density estimation method.

18. A method as claimed in claim 17 wherein the density estimation method comprises a boosting Gaussian mixture model.

19. A method as claimed in claim 1 wherein:

the confidence value is a value between zero and one; and

the confidence value corresponds to an overlap region of the predicted feature space and the normal baseline distribution such that a relatively high confidence value corresponds with a relatively large overlap region and a relatively low confidence value corresponds with a relatively small overlap region.

20. A method as claimed in claim 1 wherein:

the confidence value is a value between zero and one;

the confidence value is based on a comparison of the predicted feature space with the normal baseline distribution and the faulty baseline distribution; and

the confidence value is greater when the predicted feature space is closer to the normal baseline distribution than the faulty baseline distribution than when the predicted feature space is closer to the faulty baseline distribution than the normal baseline distribution.

21. A method as claimed in claim 1 wherein the method further comprises providing a mechanical system diagnosis indicating one or more faults.

22. A method as claimed in claim 21 wherein providing the mechanical system diagnosis further comprises inputting features into a trained self-organizing map to generate and display a health map.

23. A method as claimed in claim 22 wherein the health map comprises a plurality of regions indication a plurality of corresponding failure modes.