SOM's have been used effectively for maintenance and for diagnosing systems in the past. Using the quantization error of the test data as a metric to diagnose the system with healthy and faulty training data is not reliable in all cases as the training data might contain noise thus resulting in an unreliable result of the SOM. In this contribution we introduce a method to use SOM and U-matrix to build an undirected graph and using shortest path algorithms to define a proximity measure that represents the closeness of the test data to that of the training data thereby improving the results of the SOM more reliable for maintenance and diagnostic purposes.
展开▼
