The problem of nuclear power plant noise diagnostics was formulated as a problem of the pattern recognition theory. The use of the entropy criterion, the difference of the conditional probability density criterion, and the Karhunen-Loeve expansion for feature extraction were discussed. The Bayes’ learning was applied to decision rule development. The non-parametric K nearest neighbor method was used for the probability density estimate. These methods were applied to a boiling type and a burnout identification with the help of an acoustic noise. The acoustic noise information about the heat exchange processes was presented in the dimensionality reduced space. The Bayes’ decision rule for the burnout identification was developed. The experiments on the Universal Combined Model and the Reactor Channel Model plants have demonstrated a high efficiency of the pattern recognition theory application to the reactor noise diagnosis.