Radial basis function classifier for fault diagnostics

ABSTRACT

A realization of a program simulator for RBF neural network is proposed, which optimizes not only the weights between the hidden and the output layers, but also the positions and the widths of the kernel functions. The minimization of the number of kernel functions in hidden layer by clustering algorithm is carried out as a data pre-processing. The results from this first stage of classification are used for an initialisation of the weights between the input and the hidden layers and prevent from falling into a local minimum in the gradient procedure that evaluates the positions and the widths of the kernels in hidden layer. The effectiveness of the proposed supervised learning is demonstrated through appropriate example of a neural network application for technical diagnostics of complex objects. The task is to determine which of the vast number of possible states the examined ship engine has and to predict future faults. Having in mind the high dimensional observation vector, the significant overlapping of the fault states, the complexity from the multimode work of the engine and the noisy environment, the pattern recognition is realised by using this powerful classifying RBF network and strategies, which improve its generalization abilities. Due to improvements in training the results achieve over 90% good recognition of samples from test extract.

REFERENCES

• 1: M. T. Musavi , W. Ahmed , K. H. Chan , K. B. Faris , D. M. Hummels, On the training of radial basis function classifiers, Neural Networks, v.5 n.4, p.595-603, July-Aug. 1992 [doi>10.1016/S0893-6080(05)80038-3]
• 2: Haykin S.: "Neural networks", Mc Master University, 1994
• 3: Moody J., Darken C.: "Fast learning in network of locally- tuned processing units", Neural Computation, 1, 1989