A minicomputer-based system for reactor surveillance was developed and demonstrated on-line. The state of the reactor is characterized by the system from an analysis of noise signals, and a surveillance algorithm statistically describes normal behavior from this characterization. Hyperellipsoids are constructed from this description to enclose normal regions of behavior in the multidimensional measurement space, which represents all possible reactor states. When measurements outside normal regions are detected, the status of the reactor is suspect. Tests at the High Flux Isotope Reactor demonstrate that the algorithm can sense changes in reactor conditions that the plant instrumentation cannot detect.