Abstract
In this paper, a sensor fault detection and isolation technique is proposed using statistical methods. An enhanced reconstruction method is proposed using Singular Value Decomposition (SVD). In the traditional SVD reconstruction method, the faulty data may affect other fault free data. The enhanced SVD (ESVD) reconstruction method is a robust method to map as a normal data. The statistical hypothesis test, namely Generalized Likelihood Ratio Test (GLRT) is applied to detect the fault in the residual space. The proposed method performance is verified by the real data of Fast Breeder Test Reactor (FBTR).
1. Introduction
Continuous sensor health condition monitoring provides a variety of benefits such as improved reliability, improved safety, reduced unnecessary periodical sensor calibration testing. For monitoring and controlling application of a complex production system, a large number of distributed sensors are used to provide chronological and spatial information. However, along with the benefit of using distributed sensors, there are some risks because of the severe consequences may arise, if the signals provided by sensors are out of calibration. A faulty sensor can provide an inappropriate information that can affect the system supervision and decisions making. Therefore, continuous monitoring of the performance of the sensor, i.e., sensor fault detection and localization are important issues in current research work.
5. Conclusions
Online monitoring of the sensor physical condition can avoid many problems associated with manual calibration of the sensors. The SVD based model is developed for detection the sensor fault in Nuclear Power Plants. This paper addresses an enhanced SVD (ESVD) reconstruction method, which is superior to SVD reconstruction. It is a simple linear algebraic factorization method. The ESVD is used to generate the residual matrix by selecting few singular vectors corresponding to largest singular values. The reconstruction matrix is mapped to the normal data. The GLRT is employed in residual space to detect the faulty sensor. If the GLRT decision function crosses the threshold value, then the fault is detected. The ESVD-GLRT based fault detection method is better than PCA-GLRT and SVD-GLRT
