Estimating the remaining useful life of a power transformer based on real-time sensor data and periodic dissolved gas analyses

1. A method for estimating a remaining useful life (RUL) of a power transformer, comprising:

during a surveillance mode, iteratively performing the following operations:

receiving a set of present time-series signals gathered from sensors in the power transformer,

using an inferential model to generate estimated values for the set of present time-series signals,

performing a pairwise differencing operation between actual values and the estimated values for the set of present time-series signals to produce residuals,

performing a sequential probability ratio test (SPRT) on the residuals to produce SPRT alarms having associated tripping frequencies (TFs),

using a logistic-regression model to compute an RUL-based risk index for the power transformer based on the TFs; and

when the risk index exceeds a risk-index threshold, generating a notification indicating that the power transformer needs to be replaced; and

periodically updating the logistic-regression model based on the results of periodic dissolved gas analyses (DGAs) for the transformer to more accurately compute the RUL-based index for the power transformer.

2. The method of claim 1 , wherein during an inferential-training mode, which precedes the surveillance mode, the method further comprises:

receiving an inferential training set of time-series signals gathered from sensors in the power transformer during normal fault-free operation; and

training the inferential model to predict values of the time-series signals based on the inferential training set.

3. The method of claim 2 , wherein during an RUL-training mode, which follows the inferential training mode and precedes the surveillance mode, the method further comprises:

receiving an RUL training set of time-series signals gathered from sensors in similar power transformers while the similar power transformers are run to failure;

receiving results of periodic DGAs for the similar power transformers while the similar power transformers were run to failure;

receiving failure times for the similar power transformers after the similar power transformers were run to failure;

using the inferential model to generate estimated values for the RUL training set of time-series signals;

performing a pairwise differencing operation between actual values and the estimated values for the RUL training set of time-series signals to produce residuals;

performing a SPRT on the residuals to produce SPRT alarms having associated TFs; and

training the logistic-regression model to predict an RUL for the power transformer based on correlations among the SPRT alarm TFs, the results from the periodic DGAs for the similar power transformers, and the failure times for the similar power transformers.

4. The method of claim 1 , wherein updating the logistic-regression model based on the results of a periodic DGA involves using the results of the periodic DGA as an additional input to the logistic-regression model so that the logistic-regression model more accurately computes the RUL-based index for the power transformer.

5. The method of claim 1 , wherein the time-series signals gathered from sensors in the power transformer include signals specifying one or more of the following:

temperatures;

currents;

voltages;

resistances;

capacitances;

vibrations;

cooling system parameters; and

control signals.

6. The method of claim 1 , wherein the TFs are windowed TFs, which are associated with preceding time windows of SPRT alarms.

7. The method of claim 1 , wherein the inferential model comprises a Multivariate State Estimation Technique (MSET) model.

8. A non-transitory, computer-readable storage medium storing instructions that when executed by a computer cause the computer to perform a method for estimating a remaining useful life (RUL) of a power transformer, the method comprising:

during a surveillance mode, iteratively performing the following operations:

receiving a set of present time-series signals gathered from sensors in the power transformer,

using an inferential model to generate estimated values for the set of present time-series signals,

performing a pairwise differencing operation between actual values and the estimated values for the set of present time-series signals to produce residuals,

performing a sequential probability ratio test (SPRT) on the residuals to produce SPRT alarms having associated tripping frequencies (TFs),

using a logistic-regression model to compute an RUL-based risk index for the power transformer based on the TFs; and

when the risk index exceeds a risk-index threshold, generating a notification indicating that the power transformer needs to be replaced; and

periodically updating the logistic-regression model based on the results of periodic dissolved gas analyses (DGAs) for the transformer to more accurately compute the RUL-based index for the power transformer.

9. The non-transitory, computer-readable storage medium of claim 8 , wherein during an inferential-training mode, which precedes the surveillance mode, the method further comprises:

receiving an inferential training set of time-series signals gathered from sensors in the power transformer during normal fault-free operation; and

training the inferential model to predict values of the time-series signals based on the inferential training set.

10. The non-transitory, computer-readable storage medium of claim 9 , wherein during an RUL-training mode, which follows the inferential training mode and precedes the surveillance mode, the method further comprises:

receiving an RUL training set of time-series signals gathered from sensors in similar power transformers while the similar power transformers are run to failure;

receiving results of periodic DGAs for the similar power transformers while the similar power transformers were run to failure;

receiving failure times for the similar power transformers after the similar power transformers were run to failure;

using the inferential model to generate estimated values for the RUL training set of time-series signals;

performing a pairwise differencing operation between actual values and the estimated values for the RUL training set of time-series signals to produce residuals;

performing a SPRT on the residuals to produce SPRT alarms having associated TFs; and

training the logistic-regression model to predict an RUL for the power transformer based on correlations among the SPRT alarm TFs, the results from the periodic DGAs for the similar power transformers, and the failure times for the similar power transformers.

11. The non-transitory, computer-readable storage medium of claim 8 , wherein updating the logistic-regression model based on the results of a periodic DGA involves using the results of the periodic DGA as an additional input to the logistic-regression model so that the logistic-regression model more accurately computes the RUL-based index for the power transformer.

12. The non-transitory, computer-readable storage medium of claim 8 , wherein the time-series signals gathered from sensors in the power transformer include signals specifying one or more of the following:

temperatures;

currents;

voltages;

resistances;

capacitances;

vibrations;

cooling system parameters; and

control signals.

13. The non-transitory, computer-readable storage medium of claim 8 , wherein the TFs are windowed TFs, which are associated with preceding time windows of SPRT alarms.

14. The non-transitory, computer-readable storage medium of claim 8 , wherein the inferential model comprises a Multivariate State Estimation Technique (MSET) model.

15. A system that estimates a remaining useful life (RUL) of a power transformer, comprising:

at least one processor and at least one associated memory; and

a notification mechanism that executes on the at least one processor, wherein during a surveillance mode, the notification mechanism:

receives a set of present time-series signals gathered from sensors in the power transformer,

uses an inferential model to generate estimated values for the set of present time-series signals,

performs a pairwise differencing operation between actual values and the estimated values for the set of present time-series signals to produce residuals,

performs a sequential probability ratio test (SPRT) on the residuals to produce SPRT alarms having associated tripping frequencies (TFs),

uses a logistic-regression model to compute an RUL-based risk index for the power transformer based on the TFs; and

when the risk index exceeds a risk-index threshold, generates a notification indicating that the power transformer needs to be replaced;

wherein the notification mechanism periodically updates the logistic-regression model based on the results of periodic dissolved gas analyses (DGAs) for the transformer to more accurately compute the RUL-based index for the power transformer.

16. The system of claim 15 , wherein during an inferential-training mode, which precedes the surveillance mode, the notification mechanism:

receives an inferential training set of time-series signals gathered from sensors in the power transformer during normal fault-free operation; and

trains the inferential model to predict values of the time-series signals based on the inferential training set.

17. The system of claim 16 , wherein during an RUL-training mode, which follows the inferential training mode and precedes the surveillance mode, the notification mechanism:

receives an RUL training set of time-series signals gathered from sensors in similar power transformers while the similar power transformers were run to failure;

receives results of periodic DGAs for the similar power transformers while the similar power transformers were run to failure;

receives failure times for the similar power transformers after the similar power transformers were run to failure;

uses the inferential model to generate estimated values for the RUL training set of time-series signals;

performs a pairwise differencing operation between actual values and the estimated values for the RUL training set of time-series signals to produce residuals;

performs a SPRT on the residuals to produce SPRT alarms having associated TFs; and

trains the logistic-regression model to predict an RUL for the power transformer based on correlations among the SPRT alarm TFs, the results from the periodic DGAs for the similar power transformers, and the failure times for the similar power transformers.

18. The system of claim 15 , wherein while updating the logistic-regression model based on the results of a periodic DGA, the notification mechanism uses the results of the periodic DGA as an additional input to the logistic-regression model so that the logistic-regression model more accurately computes the RUL-based index for the power transformer.

19. The system of claim 15 , wherein the time-series signals gathered from sensors in the power transformer include signals specifying one or more of the following:

temperatures;

currents;

voltages;

resistances;

capacitances;

vibrations;

cooling system parameters; and

control signals.

20. The system of claim 15 , wherein the TFs are windowed TFs, which are associated with preceding time windows of SPRT alarms.