Techniques for confident and reliable fault detection in large scale engineering plants

Traditionally, in large scale engineering plants, fault detection is performed through the use of fixed threshold bounds. In this detection scheme, upper and lower thresholds are placed on the plant´s status data. Error flags are then produced whenever the data exceeds either of its associated bounds. The major problems with this technique are that these error flags do not produce confident indications of “true” faults, and they do not reliably temporally locate the start of the faulty behaviour. In this paper, two novel model-based techniques are presented which address these problems. The first technique is an ad hoc method directed specifically at current faults within the domain of cable amplifier networks. The second technique is a more general method based on behavioural modeling through the use of a class of asymptotically stable recurrent neural networks