Fault detection and diagnosis in batch and semi-batch processes using artificial neural networks

The objective of this work is to assess the feasibility of adopting artificial neural networks (ANNs) in fault detection and diagnosis for batch and semi-batch processes. Although there is a large volume of related publications available, most of them used steady-state data to train ANNs and, as such, the task of fault diagnosis can only be implemented in continuous operations. Based upon the concept of analytical redundancy, the framework of a two-stage fault monitoring system is proposed in this paper. In the first stage, a hybrid ANN is adopted to predict the long-term dynamic behaviors of the output variables under normal condition. The occurrence of fault(s) can be detected by inspecting the residuals, i.e. the differences between the measured and the predicted values of outputs. A second feedforward neural network is then used for the purpose of differentiating the residual patterns caused by various faults. In addition to the fact the results of pilot tests are quite satisfactory, it is also demonstrated in our experimental studies that the proposed fault-monitoring system is capable of detecting and diagnosing faults that cannot be described by traditional mathematical models.