This paper proposes a new method of failure diagnosis of nuclear power plant (NPP). Transient behavior of the NPP includes ample failure information even for a limited period of time from the failure onset. We tried to develop a diagnosis system with high capability of identifying the failure cause and of estimating failure severeness. The Walsh function transformation of transient time series data and the reduction of the Walsh coefficients into ternary valued amplitude indicators were utilized to extract the essential characteristics of failure. The correspondences of the transient characteristics and causes were summarized in a failure symptom database. A method of ternary tree search using an information measure as a heuristic strategy was adopted to conduct the efficient retrieval of failure causes in the database. Through numerical experiments using a simulation model of a NPP, the diagnostic capability of the system was proved to be satisfactory.