Information fusion of classifiers in systems with partial redundant information

• US 6,757,668 B1
• Filed: 11/03/2000
• Issued: 06/29/2004
• Est. Priority Date: 11/05/1999
• Status: Expired due to Fees

• Current Assignee: General Electric Company
• Original Assignee: General Electric Company
• Inventors: Goebel, Kai Frank, Badami, Vivek Venugopal, Ashby, Malcolm John
• Primary Examiner(s): Patel, Ramesh
• Assistant Examiner(s): Hirl, Joseph P.
• Application Number:
  US09/704,476
• Time in Patent Office: 1,334 Days
• Field of Search: 706/47, 706/59, 706/61
• US Class Current: 706/59
• CPC Class Codes: G05B 23/0221 Preprocessing measurements, e.g. data collection rate adjustment; Standardization of measurements; Time series or signal analysis, e.g. frequency analysis or wavelets; Trustworthiness of measurements; Indexes therefor; Measurements using easily measured parameters to estimate parameters difficult to measure; Virtual sensor creation; De-noising; Sensor fusion; Unconventional preprocessing inherently present in specific fault detection methods like PCA-based methods
• CPC Class Codes: G05B 23/0235 based on a comparison with predetermined threshold or range, e.g. "classical methods", carried out during normal operation; threshold adaptation or choice; when or how to compare with the threshold
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data

Abstract

The present invention provides methods and tools to aggregate information stemming from a plurality of different classification tools and supportive evidential information to arrive at a unified classification estimate. The information fusion system according to the present invention has a plurality of sensors associated with the system, where each sensor is related to at least one class of the system and is data-related to the at least one class. A plurality of classification tools are each designed to receive selected and pre-processed outputs from the sensors and to generate classification outputs representing a state of at least one class of the system. An information fusion tool is configured to receive the outputs of the classification tools as well as evidential information as inputs, and has an hierarchical architecture which manipulates the inputs to generate an output of aggregated fused information for a particular class of the system.

Claims

1. An computer implemented information fusion configuration used for generating an output state of a system, the information fusion configuration comprising:

a plurality of sensors associated with the system, each sensor related to at least one class of the system and sensing data related to the at least one class;

a plurality of classification tools each designed to receive selected outputs from the sensors, and to generate classification outputs representing a state of at least one class of the system; and

an information fusion tool configured to receive the outputs of the classification tools as inputs, and having a hierarchical architecture which manipulates the inputs to generate an output of aggregated fused information for a particular class of the system and wherein the information fusion tool uses a priori information concerning the classification tools, wherein a source of the priori information is from at least one of confusion matrices developed for individual classification tools, a relevance matrix corresponding to the classification tools supplying information to the fusion tool and a reliability matrix corresponding to the classification tools supplying information to the fusion tool.

2. The invention according to claim 1 wherein the information fusion tool includes:

a scaling layer which provides a higher weight to the output of a first classification tool having a statistically greater reliability for a specified fault than the output of a second classification tool with a statistically lower reliability.

3. The invention according to claim 2, wherein the scaling layer includes encoded a priori information derived from the reliability matrix, such that information from each classification tool being used by the information fusion tool is scaled by a degree of reliability of the classification tool.

4. The invention according to claim 2, wherein the scaling layer is configured such that when the sum of the reliabilities for a relevant class is less than a predetermined threshold value, multiplied by the relevances of the classification tools, a boost operation is undertaken for the relevant classification tools.

5. The invention according to claim 1 wherein the information fusion tool includes:

a strengthening layer which generates an output by summation of classification tool outputs which are in agreement.

6. The invention according to claim 1 wherein the information fusion tool includes:

a weakening layer which performs conflict resolution processing between at least two of the plurality of classification tools which disagree on the classification state of the system.

7. The invention according to claim 6, wherein the weakening layer includes a discount module is designed such that when classification tools disagree on a classification state of the system, the output from the classification tools are discounted.

8. The invention according to claim 6, wherein when several faults are indicated at the same time, their contribution will be discounted, for systems with a one-fault assumption.

9. The invention according to claim 1 wherein the information fusion tool includes:

an association layer designed to correlate a classification tool output by known preferred misclassifications of the classification tool.

10. The invention according to claim 9, wherein the correlation in the association layer is obtained by a pre-multiplication of a classification tool output with an association matrix. The association matrix is calculated by subtracting one from the reliability matrix where off-diagonal elements below a threshold are set to zero and where the diagonal elements are all one. The result of the calculation has a minimum cutoff of zero.

11. The invention according to claim 1 wherein the information fusion tool includes:

an evidence updating layer which supplies secondary evidentiary information which supports the determined classification state of the system; and

a tie-breaking layer designed to select between at least two equally likely classification states.

12. The invention according to claim 11, wherein, in the evidence updating layer, data from an evidence matrix is used to enforce the existence of a particular class.

13. The invention according to claim 11, wherein the tie-breaking layer includes a criticality module with information from a criticality matrix, and when several values of an intermediate fused class output are declared substantially equal, faults determined to be less critical are discounted by a predetermined value.

14. The invention according to claim 11, wherein the tie-breaking layer includes a frequency module with information from a frequency matrix, and when several values of a fused class output are declared substantially equal, faults determined to be less frequent are discounted by a predetermined value.

15. The invention according to claim 14, wherein a forgetting factor of the decision fading module is used to determine the weights applied to the later made decision and the earlier made decision.

16. The invention according to claim 1 wherein the information fusion tool includes:

a concurrency layer configured to aggregate the input information from the plurality of classification tools, and to perform an information fading operation when predefined temporal conditions exist.

17. The invention according to claim 16, wherein the concurrency layer includes a decision fading module which assigns higher weight values to a later made decision, than weight values assigned to an earlier made decision.

18. The invention according to claim 16, wherein the concurrency layer includes a decision smoothing module for use with individual classification tools, the decision smoothing module implemented via an exponential averaging time series filter with an adaptive smoothing parameter expressed as:

19. The invention according to claim 18, wherein the smoothing parameter is obtained by;

20. The invention according to claim 18, wherein the decision smoothing module is turned off when decisions from the classification tools are in disagreement.

21. The invention according to claim 1 wherein the information fusion tool includes:

a backscaling layer which manipulates the output generated by the fusion tool into a standardized output value understandable by an end user.

22. The invention according to claim 21, wherein the backscaling layer includes a backscaling module provides a scaling factor to an intermediate output value of each class, the scaling factor being based on the number of classification tools for that class;

and

wherein, a dilating operation is performed on the fused intermediate output.

23. The invention according to claim 1, wherein the information fusion tool is configured to address conflict resolution for cases,

where output of two or more classification tools disagree;

where outputs of different classification tools are generated at different times, separated from each other by more than a predetermined time period;

where classification tools are updated at different rates; and

where classification tools have varying class coverages.

24. The invention according to claim 1, wherein the classification tools are diagnostic tools, the information fusion tool is a diagnostic fusion tool, and the output of the fusion tool is a fused diagnostic output.

25. The invention according to claim 1, wherein the information fusion tool is designed to perform at least one of feature fusion, data fusion and decision fusion processings.

26. An computer implemented information fusion tool comprising:

a concurrency layer configured to aggregate input information from more than a single classification tool, and to perform an information fading operation when predefined temporal conditions exist;

an association layer configured to correlate a classification tool output by known preferred misclassifications of the classification tool;

a scaling layer configured to weigh the output of a classification tool with a statistically greater reliability for a specified fault higher than the output of a classification tool with a statistically lower reliability;

a strengthening layer configured to generate an output by summation of classification tools which are in agreement;

a weakening layer configured to perform conflict resolution processing between at least two classification tools which disagree on the classification state of the system;

an evidence updating layer configured to supply secondary non-classification tool information which is used to support a classification state;

a tie-breaking layer configured to select between at least two equally likely classification states; and

a backscaling layer configured to manipulate values generated by the fusion tool into a standardized output value.

27. The invention according to claim 26, wherein the outputs from the classification tools are manipulated sequentially by at least a sub-set of the layers of the fusion tool, until a most likely candidate class has been refined as the fused output.

28. The invention according to claim 26, wherein the fusion tool increases and decreases weights values given to the classes according to strategies implemented in at least one layer of the fusion tool.

29. An information fusion product comprising:

a computer usable medium having computer readable program code embodied therein for causing different information from more than a single classification tool to be fused into an aggregated output representing a state of a system, the computer readable program code on the computer usable medium comprising;

computer readable program code for causing a computer to aggregate input information from more than a single classification tool, and to perform an information fading operation when predefined temporal conditions exist;

computer readable program code for causing a computer to correlate a classification tool output by known preferred misclassifications of the classification tool;

computer readable program code for causing a computer to weigh the output of a classification tool with a statistically greater reliability for a specified fault higher than the output of a classification tool with a statistically lower reliability;

computer readable program code for causing a computer to generate an output by summation of classification tools which are in agreement;

computer readable program code for causing a computer to perform conflict resolution processing between at least two classification tools which disagree on the classification state of the system;

computer readable program code for causing a computer to supply secondary non-classification tool information which is used to support a classification state;

computer readable program code for causing a computer to select between at least two equally likely classification states; and

computer readable program code for causing a computer to manipulate values into a standardized output value.

30. A computer implemented method for generating a fused information output by use of an information fusion tool comprising at least one of the steps of:

aggregating input information from more than a single classification tool;

performing an information fading operation when predefined temporal conditions exist;

correlating a classification tool output by known preferred misclassifications of the classification tool;

weighing the output of a classification tool with a statistically greater reliability for a specified fault higher than the output of a classification tool with a statistically lower reliability;

generating intermediate output by summation of classification tools which are in agreement;

performing conflict resolution processing between at least two classification tools when the at least two classification tools disagree on the classification state of the system;

supplying secondary non-classification tool information to support a classification state;

selecting between at least two equally likely classification states, when the at least two equally likely classification states exist; and

manipulating values generated by the fusion tool into a standardized output value.

31. An computer implemented information fusion tool configured to receive information from a plurality of classification tools and to manipulate the received information to generate an output of fused information, the fusion tool including a combination of layers selected from a group consisting of:

a concurrency layer configured to aggregate input information from more than a single classification tool, and to perform an information fading operation when predefined temporal conditions exist;

an association layer configured to correlate a classification tool output by known preferred misclassifications of the classification tool;

a scaling layer configured to weigh the output of a classification tool with a statistically greater reliability for a specified fault higher than the output of a classification tool with a statistically lower reliability;

a strengthening layer configured to generate an output by summation of classification tools which are in agreement;

a weakening layer configured to perform conflict resolution processing between at least two classification tools which disagree on the classification state of the system;

an evidence updating layer configured to supply secondary non-classification tool information which is used to support a classification state;

a tie-breaking layer configured to select between at least two equally likely classification states; and

a backscaling layer configured to manipulate values generated by the fusion tool into a standardized output value.