An analysis of the learning, recall and generalization characteristics of neural networks for detecting and diagnosing process failures in steady state processes is presented. The single fault assumption has been relaxed to include multiple causal origins of the symptoms. The effect of incomplete and uncertain process symptom data such as sensor faults, and the effect of degradation of different hidden units, on the performance of the network, have been analyzed. Various neural network topologies (i.e. number of hidden units and hidden layers) have been tested and compared. The results show that accurate recall and generalization behavior is observed during the diagnosis of single faults. Performance during recall improves at first with an increase in the number of hidden units and with the amount of training, and then attains convergence. In general, performance during generalization improves with the extent of training. The networks are also able to diagnose correctly even in the presence of faulty operation of certain sensors. Networks trained on single faults are able to accurately diagnose measurement patterns resulting from multiple faults in a large majority of the cases studied. Graceful degradation of diagnostic function was observed in many of the multiple-fault cases that were not accurately diagnosed.