NOVELTY DETECTION AND NEURAL-NETWORK VALIDATION

One of the key factors which limits the use of neural networks in many industrial applications has been the difficulty of demonstrating that a trained network will continue to generate reliable outputs once it is in routine use. An important potential source of errors is novel in put data; that is, input data which differ significantly from the data used to train the network. The author investigates the relationship b etween the degree of novelty of input data and the corresponding relia bility of the outputs from the network. He describes a quantitative pr ocedure for assessing novelty, and demonstrates its performance by usi ng an application which involves monitoring oil flow in multiphase pip elines.