Improved on-line process fault diagnosis through information fusion in multiple neural networks

A single neural network based fault diagnosis system may not give reliable fault diagnosis due to the fact that a perfect neural network is generally difficult, if not impossible, to develop. To enhance fault diagnosis reliability, this paper proposes a technique where multiple neural networks are developed and their diagnosis results are combined to give the overall diagnosis result. To develop a diverse range of individual networks, each individual network is trained on a replication of the original training data generated through bootstrap re-sampling with replacement. Furthermore, individual networks are trained from different initial weights. Different combination schemes including averaging, majority voting, and a proposed modified majority voting are studied. Applications of the proposed method to a simulated continuous stirred tank reactor demonstrate that combining multiple neural networks can give more reliable and earlier diagnosis than a single neural network whether the networks are trained on quantitative data or qualitative trend data. It is also shown that the modified majority voting combination method proposed in this paper gives better performance than other combination schemes.