A new method of rough RBF neural network ensembles

The performance of a single neural network is limited, but multiple neural networks can achieve higher classification accuracy and efficiency than the original single classifiers. In the paper, a new method of neural network ensembles based on rough set theory is described. An extended rough set model based real-value attribute is proposed, which decides the uncertainty problem of clustering regions for RBF hidden layer units. From the rough set theory, two cluster centers, which are lower and upper approximation cluster centers, can be required. Then, under the Experience Risk Minimum criterion, the two RBF neural networks with different hidden layer units could be combined. In the end of the paper, a simulation of flight actuators fault diagnosis is given, and results show that the method is valid and effective.