Artificial neural networks in CT–PT contact detection in a PHWR

Abstract

In a pressurized heavy water reactor (PHWR), contact between calandria tubes (CT) and pressure tubes (PT) makes them susceptible to delayed hydrogen cracking. Periodic inspection of the channels must be carried out to detect the contact. As the number of channels in a PHWR is very large (306 in a 230 MW plant) periodic in-service inspection of all the channels leads to an unacceptable downtime. A non-intrusive technique that employs a system identification method is presently used for contact detection. The technique tends to overpredict the number of channels in contact, i.e. they diagnose many channels as contacting while the channels are in fact not in contact. This puts a large number of healthy channels on the at risk list reducing the efficacy of the method. This paper demonstrates the power of artificial neural networks (ANNs) in diagnosing the CT–PT contact. A counterpropagation neural network consisting of a Kohonen layer and a Grossberg layer has been employed. The noise tolerance of the network has been demonstrated.