Decomposition technique for remaining useful life prediction

• US 8,725,456 B1
• Filed: 05/05/2009
• Issued: 05/13/2014
• Est. Priority Date: 05/05/2009
• Status: Active Grant

• Current Assignee: USA As Represented By The Administrator Of The NASA
• Original Assignee: The United States of America As Represented By The Secretary of Agriculture
• Inventors: Saha, Bhaskar, Goebel, Kai F., Saxena, Abhinav, Celaya, Jose R.
• Primary Examiner(s): Breene, John
• Assistant Examiner(s): Satanovsky, Alexander
• Application Number:
  US12/454,024
• Time in Patent Office: 1,834 Days
• Field of Search: 702/34, 702/35, 702181-183
• US Class Current: 702/181
• CPC Class Codes: G05B 23/0283 Predictive maintenance, e.g. involving the monitoring of a system and, based on the monitoring results, taking decisions on the maintenance schedule of the monitored system; Estimating remaining useful life [RUL] preventive maintenance, i.e. planning maintenance according to the available resources without monitoring the system G06Q10/06

Abstract

The prognostic tool disclosed here decomposes the problem of estimating the remaining useful life (RUL) of a component or sub-system into two separate regression problems: the feature-to-damage mapping and the operational conditions-to-damage-rate mapping. These maps are initially generated in off-line mode. One or more regression algorithms are used to generate each of these maps from measurements (and features derived from these), operational conditions, and ground truth information. This decomposition technique allows for the explicit quantification and management of different sources of uncertainty present in the process. Next, the maps are used in an on-line mode where run-time data (sensor measurements and operational conditions) are used in conjunction with the maps generated in off-line mode to estimate both current damage state as well as future damage accumulation. Remaining life is computed by subtracting the instance when the extrapolated damage reaches the failure threshold from the instance when the prediction is made.

Claims

1. A program storage device embodying a program of instructions contained on non-transitory, computer readable media, executable by the computer to explicitly decompose uncertainties associated with a remaining useful life (RUL) of a worn or otherwise damaged active object, wherein the uncertainties comprise:

(i) a first uncertainty in at least one-failure precursor-feature-to-damage mapping, where the first uncertainty is numerically computed and represented as a first uncertainty distribution;

(ii) a second uncertainty in a user-defined future usage of the object, where the second uncertainty is numerically computed and represented as a second uncertainty distribution;

(iii) a third uncertainty in at least one estimated future damage growth rate, where the third uncertainty is numerically represented as a third uncertainty distribution; and

(iv) wherein the uncertainties (i), (ii) and (iii) are implemented using an uncertainty management program that numerically combines the computed first, second and third uncertainties to further compute and present to a user a remaining useful life (RUL) uncertainty for a future time, of a worn or otherwise damaged active object, as an output of a kernel function that computes a RUL, by quantifying how much time is left until functionality of the object is lost.

2. The program storage device of claim 1, wherein said uncertainty management program incorporates a Relevance Vector Machine and at least one of a Particle Filter and a Gaussian Process Regression.

3. The program storage device of claim 1, wherein said uncertainties in said RUL of said active object further comprise at least one of:

(v) a fourth uncertainty in at least one sensor value measurement, where the fourth uncertainty is numerically computed and represented as a fourth uncertainty distribution;

(vi) a fifth uncertainty in at least one failure-precursor-feature of damage to said object where the fifth uncertainty is numerically computed and represented as a fifth uncertainty distribution;

(vii) a sixth uncertainty in at least one estimated future operating condition, where the sixth uncertainty is numerically computed and represented as a sixth uncertainty distribution;

(viii) a seventh uncertainty in at least one estimated future sensor value, where the seventh uncertainty is numerically represented as a seventh uncertainty distribution; and

(ix) at least one correlation between a past usage of the object and at least one uncertainty in said user-defined future usage of said object, where the correlation is numerically computed and represented as an eighth uncertainty distribution;

(x) wherein at least one of the fourth, fifth, sixth, seventh and eighth uncertainties is implemented using an uncertainty management program and is numerically combined with said computed first, second and third uncertainties to further compute and present to a user said remaining useful life (RUL) uncertainty for a future time, of said worn or otherwise damaged active object, as an output of a kernel function that computes a RUL, by quantifying how much time is left until functionality of the object is lost.

4. The program storage device of claim 3, wherein said uncertainty management program incorporates a Relevance Vector Machine and at least one of a Particle Filter and a Gaussian Process Regression.

5. A program storage device embodying a program of instructions contained on non-transitory, computer readable media, executable by a computer to predict or estimate remaining useful life (RUL) of a worn or otherwise damaged active object from at least one characteristic of damage to the object, the instructions comprising:

(i) functional decomposition of a damage progression learning task, for a worn or otherwise damaged active object, into at least two independent parts comprising current damage estimation and damage growth rate estimation;

(ii) measurement and collection of training data, comprising sensor measurements, operating conditions, and at least one ground truth damage attribute;

(iii) identification of at least one precursor of failure feature of the object;

(iv) provision of a first mapping of failure-precursor-feature-to-damage that associates a precursor of failure feature of the object with current damage to the object;

(v) provision of a second mapping that associates at least one operating condition for the object with growth rate of damage to the object;

(vi) measurement and collection of run-time data from the object, including at least one present operating condition and at least one estimated future operating condition for the object, relation of each of at least one of the failure precursor features to current damage to the object using the first mapping and at least one ground truth damage attribute for the object;

(vii) identification of at least one failure precursor feature of the object from the run time data;

(viii) use of the first mapping to estimate the current damage for the object;

(ix) use of the second mapping of the at least one operating condition with the growth rate of damage to the object to estimate a future damage growth rate from at the least one future operating condition;

(x) provision of a selected failure threshold and extrapolation of the damage growth rate to a failure threshold; and

(xi) computation and presentation of a remaining useful life (RUL) for the object, measured as a difference between estimated time when the failure threshold will be reached and a present time;

(xii) provision of a first uncertainty in said at least one failure-precursor-feature-to-damage mapping, where the first uncertainty is numerically computed and represented as a first uncertainty distribution;

(xiii) provision of a second uncertainty in a user-defined future usage of said object, where the second uncertainty is numerically computed and represented as a second uncertainty distribution;

(xiv) provision of a third uncertainty in said at least one estimated future damage growth rate, where the third uncertainty is numerically represented as a third uncertainty distribution; and

(xv) implementation of the uncertainties (xii), (xiii) and (xiv) using an uncertainty management program that numerically combines the computed first, second and third uncertainties to further compute and present to a user a remaining useful life (RUL) uncertainty for a future time, of said worn or otherwise damaged active object, as an output of a kernel function that computes a RUL, by quantifying how much time is left until functionality of the object is lost.

6. The program storage device of claim 5, wherein said estimated future damage growth rate is correlated with at least one damage growth mapping, chosen to correspond to changes in said at least one ground truth damage attribute.

7. The program storage device of claim 6, wherein said estimated future damage growth mapping is represented as a nonlinear damage growth model, with at least one nonlinear growth model parameter chosen to correspond to said at least one ground truth damage attribute.

8. The program storage device of claim 5, wherein a number of identified failure precursor features in a set is reduced by discarding at least one of said failure precursor features that (i) is already included in the identified set of failure precursor features that are associated with said damage to said object and (ii) is correlated with at least one other failure precursor feature in the identified failure precursor feature set with a correlation value that is at least equal to a threshold correlation value.

9. The program storage device of claim 5, wherein said ground truth damage attribute for said current damage is obtained through direct observation of said object.

10. The program storage device of claim 5, wherein said ground truth attribute for said current damage is obtained from an observation that is not a direct measure of said current damage of said object.

11. The program storage device of claim 5, wherein said ground truth attribute for said current damage is obtained from collected training data where said object is operated and training data are collected until said current damage exceeds said failure threshold.

12. The program storage device of claim 5, wherein said uncertainty management program incorporates a Relevance Vector Machine and at least one of a Particle Filter and a Gaussian Process Regression.

13. A program storage system embodying a program of instructions contained on non-transitory, computer readable media, executable by the computer, for predicting or estimating remaining useful life (RUL) of a worn or otherwise damaged active object from at least one characteristic of damage to the object, the system comprising:

(i) a problem formulation component that functionally decomposes the damage progression learning task into at least two independent parts comprising (i) accumulated damage estimation, and (ii) damage growth rate estimation;

(ii) a data measurement and collection component that measures and collects training data, comprising sensor measurements, object operating conditions, and at least one ground truth damage attribute;

(iii) a failure precursor feature extraction component that identifies at least one precursor of failure feature of the object;

(iv) a first data analysis component that provides a first mapping that associates a failure precursor feature of the object with accumulated damage to the object;

(v) a second data analysis component that creates a second mapping that associates at least one operating condition for the object with growth rate of damage to the object;

(vi) a run time data collection component that measures and collects run time data from the object, including at least one present operating condition and at least one computed future operating condition for the object, relation of each of at least one of the failure precursor features to damage to the object, and at least one ground truth damage attribute for the object;

(vii) a failure precursor feature extraction component that identifies at least one failure precursor features for the object from the run time data;

(viii) a damage estimation component that uses the first mapping to estimate current damage to the object;

(ix) a damage prediction component that;

(a) uses the second mapping of operating conditions with object damage growth rate to estimate a future damage growth rate from at least one future operating condition;

(b) provides a failure threshold and extrapolates the computed damage growth rate to the failure threshold; and

(c) computes and presents a remaining useful life (RUL) for the worn or otherwise damaged active object, measured as a difference between estimated time when the failure threshold will be reached and a present time; and

(x) an uncertainty management component that numerically computes uncertainties in (a) a failure-precursor-feature-to-damage mapping for the object, (b) a user-defined future usage of the object, and (c) a future damage growth rate, and combines computed uncertainties in (a), (b) and (c) to further compute and present to a user a remaining useful life (RUL) uncertainty for a future time, of the worn or otherwise damaged active object, as an output of a kernel function, by quantifying how much time is left until functionality of the object is lost.

14. The system of claim 13, wherein said damage growth is further represented with at least one damage growth mapping, chosen to correspond to said at least one ground truth damage attribute.

15. The system of claim 14, wherein said damage growth mapping is further represented as a nonlinear damage growth model, with at least one nonlinear growth model parameter chosen to correspond to said at least one ground truth damage attribute.

16. The system of claim 13, wherein an identified failure precursor feature set is reduced in size by discarding at least one of said failure precursor features that (i) is already included in said identified set of failure precursor features that are associated with said damage to said object and (ii) is correlated to at least one other failure precursor feature in the identified failure precursor feature set with a correlation value at least equal to a selected threshold correlation value.

17. The system of claim 13, wherein said ground truth attribute is obtained for said current damage through direct observation.

18. The system of claim 13, wherein said ground truth attribute is obtained for said current damage by derivation from an observation that is not a direct measure of said current damage condition.

19. The system of claim 13, wherein said ground truth attribute is obtained for said current damage from collected training data where said object is operated and training data are collected until said damage exceeds said failure threshold.

20. The system of claim 13, wherein said uncertainty management component incorporates a Relevance vector Machine and at least one of a Particle Filter and a Gaussian Process Regression.

21. The system of claim 13, wherein-said uncertainty management component further numerically computes an uncertainty in at least one of:

(d) present sensor value measurements, (e) said failure precursor feature computations, (f) future operating conditions, (g) correlation between a past usage of said object and said future usage of the said object, and (h) a future sensor value measurement, and combines computed uncertainties, in addition to said uncertainties in (a), (b) and (c) in claim 13, to further compute and present to a user said remaining useful life (RUL) uncertainty for a future time, of said worn or otherwise damaged active object, as an output of a kernel function that computes a RUL, by quantifying how much time is left until functionality of said object is lost.

22. The system of claim 21, wherein said uncertainty management component incorporates a Relevance Vector Machine and at least one of a Particle Filter and a Gaussian Process Regression.