Sensor fault detection in nuclear power plants using multivariate state estimation technique and support vector machines.

Recent developments in artificial intelligence at Argonne National Laboratory (ANL) have culminated in the capability to perform nuclear power plant sensor validation and early fault detection in an integrated package called the Multivariate State Estimation Technique (MSET). Nuclear reactor signals are validated by comparing signal prototypes with the actual reactor signals. Residuals from these comparisons are used in a sensitive hypothesis testing method, the Sequential Probability Ratio Test (SPRT). The SPRT examines the stochastic components of the residuals and can detect if the statistical characteristics begin to change. The signal prototypes are estimated based on empirical data. The property of an estimation algorithm to make predictions on limited amount of data is designated as generalization ability. It is a very important issue in algorithm selection. Recently, we included a new machine learning algorithm called the Support Vector Machines (SVM) in the estimation module of MSET. In the SVM algorithm, the input data space (set of reactor signals) is transformed into a high-dimensional nonlinear space using a kernel function, and the learning problem is formulated as a convex quadratic programming problem with a unique solution. In particular, we implemented and tested several kernels developed at Argonne National Laboratory. Our recent results indicated that the combination of MSET kernels with the SVM method has better noise reduction and generalization properties than the standard MSET algorithm. In this paper we compare fault detection properties of these algorithms. “Present address: Sun Microsystems, 901 San Antonio Road, Palo Alto, CA 94303, USA NOV08 ZM)