抄録
This paper concerns the retraining of learning support vector machines (SVMs) that are trained to classify abnormal operating conditions of mechanical systems. Detection of the emergence of new data classes is mainly discussed that can trigger reconstruction of the training data set and retraining of the SVMs to obtain updated discriminants. In our approach, two self-organizing maps (SOMs), one with short term memory representing the current tendency and the other with long term memory representing the acquired knowledge, are used to detect the changes in the data structure by mapping the reference vectors in the long term SOM onto those in the short term SOM. A simple example using the data collected from draw-texturing machines is provided.
