Improved on-line process fault diagnosis using stacked neural networks

Since it is generally difficult, if not impossible, to develop a perfect neural network, a single neural network can lack reliability. Therefore a single neural network based fault diagnosis system may not give reliable fault diagnosis. Neural network model reliability or robustness can be improved by combining several non-perfect neural networks. Each individual network is trained on a bootstrap re-sample of the original training data. The outputs from the individual networks are averaged to give the final diagnosis results. Applications of the proposed method to a continuous stirred tank reactor demonstrate that a stacked neural network can give more reliable diagnosis than a single neural network.