Semi-supervised classification method for dynamic applications

The behaviour of a dynamic system can assume different operating states in the course of time. In pattern recognition methods, each state is represented by a set of similar patterns forming restricted regions in the feature space, called classes. Recognising the state, class, of a new incoming pattern can be performed using a membership function. The membership function accuracy depends on the prior knowledge about the system functioning. For dynamic systems, this knowledge often suffers from two drawbacks. Firstly, there is no prior information about some states. Thus, the occurrence of new states must be detected and integrated online in the data set. Secondly, the prior information about some states, especially the faulty ones, is not sufficient to properly estimate their membership functions. The missing information can be obtained from the new classified patterns. Thus, these membership functions must be adapted online with the classification of new incoming patterns. In this paper, we propose a semi-supervised classification method based on fuzzy pattern matching (FPM). The goal is to learn membership functions with a limited initial data set. The class membership function according to each feature is sequentially learned with the occurrence of patterns and then it is updated online using an incremental or recursive approach. This learning method does not require any prior information about the nature of classes or their number.