New bad-data detection and identification technique based on rotation of measurement order for sequential state estimation

Abstract

The paper describes the development of a new bad-data detection and identification technique for sequential state estimations using Givens orthogonal transformation. In the sequential state estimation, measurements are divided into basic measurements and redundant ones. Basic measurements are processed first, and then redundant ones are processed one at a time. The successive change in the cost quadratic function, denoted by δJ, resulting from the processing of a redundant measurement, follows a Chi-square distribution with one degree of freedom. If there are no gross errors in the measurements processed up to the present point, then δJ is free from the contamination of gross errors in the measurements which are waiting to be processed. Furthermore, δJ shows the degree of consistency between the present measurement and those processed before. Hence, the bad-data detection by δJ test is performed, after one measurement has been processed, by comparing the value of δJ with a constant obtained from the Chi-square distribution. The new bad-data identification method is based on the fact that despite the existence of a few gross errors, most of the measurements are ‘good’ in the sense that they are corrupted by Gaussian noise only. Thus, by searching for these ‘good’ ones and processing them first, suspect measurements are isolated to the bottom of the measurement list. This is done by successively rotating the order of measurements while they are being processed. The proposed bad-data detection and identification method does not rely on the analysis of measurement residuals. The new bad-data processing technique is tested on an industrial steam turbine system for which a robust and efficient state estimator has been developed. Numerical results for bad-data detection and identification tests are reported.