Integrating Unsupervised and Supervised Learning in Neural Networks for Fault Diagnosis

Recently, there has been considerable interest in the use of neural networks for fault diagnosis applications. To overcome the main limitations of the neural networks approach, improvements are sought mainly in two respects: (a) a better understanding of the nature of decision boundaries (b) determining the network structure without the usual arbitrary trial and error schemes. In this perspective, we have compared different neural network paradigms and developed an appropriate integrated approach. A feedforward network with ellipsoidal units has been shown to be superior to other architectures. Two different types of learning strategies are compared for training neural networks: unsupervised and supervised learning. Their relative merits and demerits are discussed and a combination has been proposed to develop a network that meets our diagnosis requirements. Unsupervised learning component serves to identify the features and establish the network structure. Supervised learning serves to finetune the resulting network. We present results from a reactor-distillation column case study to demonstrate the structure of the measurement pattern distribution and the suitability of ellipsoidal units approach. By considering the transient behavior in the diagnosis framework, we point out that the problem of fault diagnosis can be treated on the same footing for both batch and continuous processes.