Data processing: artificial intelligence – Knowledge processing system – Creation or modification
Reexamination Certificate
2000-11-03
2004-06-29
Patel, Ramesh (Department: 2121)
Data processing: artificial intelligence
Knowledge processing system
Creation or modification
C706S047000, C706S061000
Reexamination Certificate
active
06757668
ABSTRACT:
FIELD OF THE INVENTION
The present application relates generally to information aggregation and more particularly to methods and tools used to aggregate at least partially redundant information developed by different classification tools, also called classifiers, to arrive at a unified estimate concerning the state of a system.
BACKGROUND OF THE INVENTION
The field of classification encompasses a wide range of disciplines including fault diagnosis, pattern recognition, modeling, and clustering, among others. A classification tool is used to analyze data through categorization into different classes. This may be accomplished using generalizations over the records of a training data set, which distinguish predefined classes to predict the classes of other unclassified records.
An issue that can make classification a difficult task is the dynamic nature of many systems. As a system signature changes, it escapes the carefully partitioned regions of a specific classification tool. Efforts such as adaptive classification attempt to cope with this challenge. An added level of complexity comes through the integration and aggregation of redundant information which has to be ranked based on its integrity.
In the case of dynamic systems, the aggregation scheme has to deal with several issues including temporally disjointed information, which is information produced at different sampling frequencies. For example, some classification tools may operate at a millisecond sample period while others give only one estimate at a specific phase of system operation. This situation is a problem when a significant event occurs between the estimate of one classification tool and the estimate of another classification tool.
To illustrate, consider two classification tools A and B. At a base case both tools render their information at the same time. Tool A can only establish whether or not event x is present while tool B can detect events x and y. Tool A can identify event x very reliably while tool B's performance is only mediocre. If event x occurs, there is a high likelihood that tool A indicates event x and also a lesser likelihood that tool B indicates event x. If both tools indicate event x there is reasonable certainty that event x did indeed occur. If tool B indicates event y one might be tempted to choose event x because of the good performance of tool A. However, should event y occur—of which tool A knows nothing about—tool A will always misdiagnose that event as either event x or a nil event. Tool B will perform its mediocre job by sometimes indicating event y correctly. The decision makers (maintenance personnel, fleet management, etc.) faces the quandary which tool to believe. They only see that (otherwise reliable) tool A indicates event x (or nil) while tool B indicates event y (or worse, since it is not highly reliable, sometimes event x). Incorrect classification is the likely result with traditional averaging and voting schemes.
Now assume that first tool A carries out its job, then tool B. Consider that first event x is present and event y occurs after tool A has completed its classification task but before tool B. An added problem is that both tools could perform correctly and that an a priori discounting of tool B's estimate is not appropriate. Rather, the temporal nature of the information has to be taken into account.
Today, service providers and manufacturers develop different classification tools to accommodate particular classification needs. These different tools are sometimes employed either for successive implementation to achieve better class coverage, or to obtain a better overall classification result. Classification tools are deployed in this manner because any one tool cannot deal with all classes of interest at the desired level of accuracy. For example, certain classification tools respond to environmental changes with less deterioration of their classification capabilities than others; some cannot easily be expanded to include new classes of investigation; some have a high reliability at detecting certain classes of faults but are not reliable for others. This patchwork approach achieves optimization at the component level, but ignores benefits gained by taking a system-level view.
In order to obtain this system-level view, information fusion technology gathers and combines the results of different classification tools to maximize the advantages of each individual tool while at the same time minimizing the disadvantages. Fusion schemes strive to deliver a result that is better than the best result possible by any one tool used. In part this can be accomplished because redundant information is available, and when correctly combined improves the estimate of the better tool and compensates for the shortcomings of the less capable tool.
Information fusion tools typically do not know whether an event occurs. As far as such tools are concerned, either one estimate is bad (and it does not know a priori which one that is) or the tools are right and something has happened. But even in a static case, that is, when classification tools provide their opinion to a fusion tool at the same time, there are a number of issues which need to be addressed. The trivial case is when several classification tools agree on the state of a class. In this situation the resulting output carries a high degree of confidence. However, if the classification tools disagree, a decision must be made as to which tool to believe or to what degree. In addition, information may be expressed in different domains, such as but not limited to probabilistic information, fuzzy information, binary information, etc. The fusion scheme needs to map the different domains into a single common domain to properly use the data. The fusion scheme also has to deal with tools that are operated at different sampling frequencies.
The majority of existing information fusion tools and techniques are configured using known classification tools. Essentially, these fusion tools perform a classification of the classification tools. For example, raw data is provided to a plurality of classification tools, and the outputs of these classification tools are in turn provided to a single classification tool which functions as the “fusion tool”. The fusion tool uses the outputs from the classification tools as an input pattern for its operation. Thus, the fusion tool simply receives the output from plurality of classification tools as an input pattern, and based on the input pattern, determines a certain classification. Such a system is classifying the received combination of outputs from the plurality of classification tools to represent a particular class. This design simply elevates the classification process away from the raw data by adding an interim process between the raw data (generated, for example, by known sensors) and the fusion tool (i.e., a known classification tool being used as a fusion tool).
Information fusion is a relatively new field, with many areas of opportunity uninvestigated, including the use of information fusion for the diagnosis of faults in a system.
In consideration of the foregoing, it has been determined desirable to develop a fusion tool that gathers the output of different classification tools and leverages the advantages of each of the outputs. Such a tool should take into consideration that individual classification tools may provide information in different domains, at different operational frequencies, have different reliability, provide different fault coverage, and may generate results which conflict with other tools. It will also be useful for the fusion tool to take into consideration the operational regimes of a system being investigated, and the need to consider secondary information related to the system.
SUMMARY OF THE INVENTION
The present invention provides methods and tools to aggregate information stemming from different classification tools and supportive evidential information to arrive at a unified classification estimate. The information fusion system includes a plurality o
Ashby Malcolm John
Badami Vivek Venugopal
Goebel Kai Frank
General Electric Company
Goldman David C.
Hirl Joseph P.
Patel Ramesh
Patnode Patrick K.
LandOfFree
Information fusion of classifiers in systems with partial... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Information fusion of classifiers in systems with partial..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Information fusion of classifiers in systems with partial... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3329054