Some effects of database corruption in system prediction performance

Many types of intelligent adaptive systems use vast databases of a-priori knowledge during training phases. These systems are then reliant on both the accuracy of this data and on the breadth of the data. It is assumed whilst training that the data encompasses the total operating window for the system in enough detail to generate an accurate ‘black box’ model of the plant under control. It may be that under certain unforeseen operating conditions, or in a scenario where there is little prior knowledge, the system may be forced to operate outside the scope of the original a-priori knowledge. Lastly the data gathered into the a-priori source may have been unintentionally corrupted. This paper aims to examine some of these effects upon two common adaptive intelligent tools, neural networks and an adaptive neuro-fuzzy inference system, ANFIS, network.