Integrated diagnosis using information-gain-weighted radial basis function neural networks

A new approach, the information-gain-weighted radial basis function neural network (RBFNN), has been proposed for machinery diagnosis in a manufacturing environment. This approach is based on the composite neural network, in which a series of RBFNNs are integrated together to perform the task of classification. Each RBFNN has only one output node and is treated as a sub-network. Unlike the conventional RBFNNs, in which only the outputs of hidden nodes are weighted, the information-gain-weighted RBFNN uses a weighting vector also in its input layer. In addition, the weighting vector is obtained according to the information gain of each input index to the process of diagnosis. In the diagnosis strategy, one sub-net is responsible for diagnosing one specific fault, while the composite network as a whole can diagnose different kinds of faults. By this approach, classification among known faults can be made and a novel fault, if any, can also be identified. This approach has been tested in the diagnosis of an internal thread tapping process. The results showed that the information-gain-weighted RBFNN can produce better distinction between conditions and the scheme of a composite neural network is indeed an improved structure for machinery diagnosis in the manufacturing environment.