Model-diversity technique for improved proactive fault monitoring

1. A method for predicting a value of a signal from a sensor, comprising:

generating N models to predict the value of the signal based on a set of quantized telemetry signals, wherein a given model produces a value of the signal using a subset of the set of quantized telemetry signals, and wherein the subset is selected from the set of quantized telemetry signals based on an objective criterion; and

predicting the value of the signal by aggregating the values produced by the N models.

2. The method of claim 1 , wherein the objective criterion accounts for a robustness of the predicted value when degradation occurs in one or more sensors in the subset.

3. The method of claim 1 , wherein the objective criterion accounts for spillover, which determines how much a disturbance on a given signal included in the subset of the quantized telemetry signals used in the given model affects the predicted value of the signal.

4. The method of claim 1 , wherein the objective criterion includes an error between the predicted value and an actual value of the signal.

5. The method of claim 4 , wherein the error includes a mean-square error.

6. The method of claim 1 , wherein the objective criterion includes cross-correlation coefficients between the quantized telemetry signals.

7. The method of claim 1 , wherein the aggregating of the values includes averaging the values.

8. The method of claim 7 , wherein, during the averaging, the values are weighted based on the variance of the values.

9. The method of claim 8 , wherein the weighting is determined using a Kalman filter.

10. The method of claim 1 , wherein generating the given model includes:

initializing a pool of candidate models which include at least two candidate models, wherein each candidate model in the pool includes a subset of the set of quantized telemetry signals; and

optimizing the given model by iteratively,

selecting two models U and V from the pool of candidate models, wherein models U and V best predict the signal based on training data in the set of quantized telemetry signals and the objective criterion;

using a genetic technique to create an offspring model W from U and V by combining parts of the two models U and V; and

adding W to the pool of candidate models.

11. The method of claim 10 , wherein each candidate model in the pool is represented using an n-bit binary string; and

wherein creating the offspring model W from U and V involves:

selecting a crossover point k, wherein k is an integer between 1 and n;

creating W by combining U and V such that the first k bits of W are the first k bits of U, and the last (n−k) bits of W are the last (n−k) bits of V; and

reversing one bit within W with a predetermined probability p.

12. The method of claim 11 , wherein k is randomly or pseudo-randomly selected.

13. The method of claim 1 , further comprising using the estimated value to predict faults in a computer system prior to failure of the computer system.

14. The method of claim 1 , further comprising monitoring a set of sensors to obtain the quantized telemetry signals.

15. A computer program product for use in conjunction with a computer system, the computer program product comprising a computer-readable storage medium and a computer-program mechanism embedded therein for predicting a value of a signal from a sensor, the computer-program mechanism including:

instructions for generating N models to predict the value of the signal based on a set of quantized telemetry signals, wherein a given model produces a value of the signal using a subset of the set of quantized telemetry signals, and wherein the subset is selected from the set of quantized telemetry signals based on an objective criterion; and

instructions for predicting the value of the signal by aggregating the values produced by the N models.

16. The computer-program product of claim 15 , wherein, during the averaging, the values are weighted based on the variance of the values.

17. The computer-program product of claim 16 , wherein the weighting is determined using a Kalman filter.

18. The computer-program product of claim 15 , wherein generating the given model includes:

initializing a pool of candidate models which include at least two candidate models, wherein each candidate model in the pool includes a subset of the set of quantized telemetry signals; and

optimizing the given model by iteratively,

selecting two models U and V from the pool of candidate models, wherein models U and V best predict the signal based on training data in the set of quantized telemetry signals and the objective criterion;

using a genetic technique to create an offspring model W from U and V by combining parts of the two models U and V; and

adding W to the pool of candidate models.

19. The computer-program product of claim 18 , wherein each candidate model in the pool is represented using an n-bit binary string; and

wherein creating the offspring model W from U and V involves:

selecting a crossover point k, wherein k is an integer between 1 and n;

creating W by combining U and V such that the first k bits of W are the first k bits of U, and the last (n−k) bits of W are the last (n−k) bits of V; and

reversing one bit within W with a predetermined probability p.

20. An apparatus, comprising:

a set of sensors configured to produce quantized telemetry signals;

a processor;

memory configured to store the quantized telemetry signals produced by the set of sensors; and

a program module, wherein the program module is stored in the memory and configured to be executed by the processor, the program module including:

instructions for generating N models to predict a value of a signal based on the set of quantized telemetry signals, wherein a given model produces a value of the signal using a subset of the set of quantized telemetry signals, and wherein the subset is selected from the set of quantized telemetry signals based on an objective criterion; and

instructions for predicting the value of the signal by aggregating the values produced by the N models.