On-line classification of acoustic burst signals by a neural network application to loose parts monitoring in nuclear power plants

Acoustic signature analysis is increasingly being used as a tool for online assessing the mechanical integrity of components in the primary circuit of nuclear power plants. During operation, the acoustic signals of loose parts monitoring system sensors are continuously monitored for signal bursts associated with metallic impacts. With the availability of neural networks new powerful tools for classification and diagnosis of burst signals can be realized online. Since signals of same event type can have similar but diverse signal forms according to their random flow-induced excitation, the characterization potential of neural networks has been used for type classification. A backpropagation neural network based on five precalculated signal parameter values has been set up for identification of three different signal types. In a pilot application at a plant, the acoustic burst signals at a steam generator were automatically monitored, classified and trended. The paper presents the successful results of six weeks online signal classification at the plant