A high noise recognition system with self-organizing features

On the basis of the knowledge of software fault tolerance, a new N-version programming scheme to pattern recognition is presented. Each architecture of the N-version scheme includes the training and recall stage. In the training stage, the training vectors are separated into several subset vectors (regions). Subsequently, the authors use the cluster discovery network to extract the features of each subset vector and concurrently complete the feature links to build the architecture for recall stage. In the recall stage, the recognition is based on the feature links between the feature layer and recognition layer. Every pattern node in the recognition layer is fed to the measure similarity block to determine the final output. Then, the authors construct a single architecture and the N-version programming based on the different region definitions of input vector for the purpose of fault tolerance. The authors´ experiments are based on the 3-version programming scheme and the recognition rate is very high for the high noise injection. In particular, the authors present a recognition radar map to measure the ability of fault tolerance. Finally, the training and recall time of the proposed system are very short under the single architecture. Therefore, the system may be used in the real-time system