Dynamic supervised classification method for online monitoring in non-stationary environments

The monitoring of a system functioning is achieved using a classifier which determines at each instant the class of a new incoming pattern. In non-stationary environments, the classifier must be able to adjust its parameters according to changes in the environment conditions. This requires a continuous learning while new patterns are available. Incremental learning is an efficient continuous learning technique for updating the classifier parameters without starting from scratch every time a new pattern is available. However in non-stationary environments, data characteristics may drift over time. This leads to deteriorate dramatically the performance of incremental learning algorithms over time. This is due to the use of data which is no more consistent with the characteristics of new incoming data. Thus, a mechanism to use only the recent and representative patterns to update the classifier parameters without a “catastrophic forgetting” is necessary. In this paper, we propose a dynamic pattern recognition method, named Dynamic Fuzzy Pattern Matching, to be used for the online monitoring of non-stationary processes functioning. This method is based on the use of an incremental algorithm allowing to follow the accumulated gradual changes of classes characteristics after the classification of each new pattern. When the accumulated gradual changes reach a suitable predefined threshold, the classifier parameters are adapted online using the recent and useful patterns.