Improving the performance of CMFD applications using multiple classifiers and a fusion framework

Abstract

In this paper, we demonstrate that the performance of a condition monitoring and fault diagnosis (CMFD) system may be improved by combining the outputs from three ‘primary’ classifiers using a novel, hybrid, data-fusion approach. The resulting classifier system involves four key processing stages. In Stage One, three diverse primary classifiers are employed. In Stage Two, the outputs from the primary classifiers are combined using majority voting. In Stage Three, any unclassified patterns from Stage Two are reassessed using Dempster-Shafer theory. Finally, in Stage Four, a simple rule base (‘expert system’) is used to integrate the results from the earlier stages. We demonstrate the effectiveness of this framework on a data set intended to allow the detection of static thermostatic valve faults in a diesel engine cooling system. Overall performance of the classifier system is shown to improve from approximately 89% (using the best of the primary classifiers) to approximately 99% (using the framework). In addition, the misclassification level of the original primary classifiers is shown to be approximately 10%, while the equivalent rate for the fusion classifier is approximately 1%. We go on to describe the application of the same classifier framework to a very different problem domain (medical diagnostics) and obtain a similar improvement in system performance. Here, overall performance of the classifier system is again shown to improve from approximately 87% (using the best of the primary classifiers) to approximately 99% (using the framework). In this case, the misclassification level of the original primary classifiers is approximately 12%, while the equivalent rate for the fusion classifier is approximately 1%. The results suggest that this framework may be appropriate for use in a range of application areas.