An advanced diagnostic method is proposed that uses automated pattern recognition for reactor noise. The method enables intensive diagnosis of known anomalies and extensive detection of unknown plant states. It also enables automatic learning of reference noise patterns for an unknown plant state and monitoring of the subsequent state change by regarding the new reference patterns as those for a known plant state. Application results for the method used on artificial noise data produced by a fast breeder reactor noise simulator are presented. A diagnostic system based on the proposed method will make it possible to automatically accumulate and make the most of anomaly data from actual power plants, although it is still difficult to identify the cause of an abnormality automatically.