Application of neural networks for check valve monitoring

A novel approach is presented for identifying possible malfunctions of check valves operating in Nuclear Power Plants (NPPs). The technique is based on the utilization of Artificial Neural Networks (ANNs) as means for classifying acoustic signatures obtained from accelerometers mounted on different locations of a check valve. The spectra computed from the time-signatures are presented to two types of pre-trained neural networks. In the first case, data sets corresponding to normal check valve operating conditions are used for training a backpropagation neural network establishing the reference model for satisfactory valve performance. The degree of mismatch between the network reference model and the operating valve responses during the recall process are used for classifying different valve conditions. In the second case, self-organizing maps (SOMs) were employed for clustering signal spectra with similar characteristics. Spectra corresponding to normal operating conditions are clustered separately from those identifying malfunctioning valves, making feasible the classification of separate check valve conditions. Both of these techniques show promise as a means of non-invasively identifying failure or malfunctioning of a check valve.